Forecasting the potential efects of significant events is an important task that can help inform proactive decision making. Using interpretable and explainable models for this challenging task could be greatly beneficial for developing trust in the predictions made by the system. In this demo, we present a prototype event prediction system using the EvCBR model, which uses a case-based reasoning approach to make predictions about cause-efect event pairs in a Knowledge Graph (KG). By performing reasoning over past cases of causal event relations present in the KG, EvCBR naturally provides explainability to its prediction results while also enabling easy integration of existing open-source KGs. Importantly, EvCBR is designed to perform inductive link prediction on unseen events and requires no training, allowing it to overcome the common limitations of other KG completion models. We demonstrate how EvCBR can be used together with a portion of the Wikidata KG to query for relevant entities, make predictions about ongoing events, and visualize past event cases and reasoning paths through the KG that were used to make predictions for the ongoing events. Our code is available at https://doi.org/10.5281/zenodo.8341874.
CEUR ceur-ws.org
Being able to predict the efects of ongoing events is an important endeavor, enabling key decision-makers to gain greater insights about potential future outcomes and thereby proactively take measures to avail of opportunities or mitigate damage. As a simple example, one may predict that an earthquake event may cause a tsunami event, which subsequently may be used to issue evacuation warnings to local residents. As more variables become involved with the events, it could also become more dificult to produce efective predictions – for example, whether an earthquake leads to a tsunami will depend on the location where the event actually took place. Making such predictions about the consequences of events is of interest to various domains and use cases, such as disaster relief planning and financial analysis.
While making predictions about events with less straightforward causes and consequences, it
is important for a model to be able to explain why it made certain predictions. This is valuable
nEvelop-O
(O. Hassanzadeh)
∗Part of this work was done while working at IBM Research.
CEUR
Workshop
Proceedings
to convince users why a prediction may be reasonable, as well as to understand situations under
which the predictions made by the model should not be taken at face value. A knowledge-driven
approach to event prediction could therefore be beneficial, since methods leveraging background
knowledge often tend to be more accurate but also more interpretable than many black-box
models. While the task of event prediction has a rich history across many domains [
To tackle this challenge, we utilize a case-based reasoning model, EvCBR [
Compared to past work, EvCBR is not limited to only making predictions about entities seen
during training, as is the case for many embedding-based prediction methods over KGs [
The core idea in our work is to frame event prediction using a KG as a 2-hop link prediction task, where we aim to predict properties about some unseen efect event that is caused by a cause event. To provide greater explainability and to support predictions for new cause events, we develop a case-based reasoning approach to find similar past events and retrieve relevant patterns to make predictions.
Intuitively, case-based reasoning techniques aim to solve new problems using past cases of
similar problems and their solutions. Extending this idea to event prediction and data present in
KGs, we treat events in the KG connected by causal relations [
A high-level overview of EvCBR’s methodology is shown in Figure 1. Starting from a new
event, such as a Protest in Iran, we proceed with making predictions about a new efect event
that may be caused by it. This example is an illustration of our attempt to predict consequences
of the protests in Iran [
Retrieving Cause-Efect Event Cases: For a given input event for which we wish to make predictions, we first identify cases of cause-efect event pairs in the KG that are similar to our new input event. We select cases based on the similarity of the case’s cause to our new query event as well as the similarity of the case’s efect’s properties to the properties about our new, unseen efect event for which we wish to make predictions.
The similarity of individual entities in the KG are computed based on outgoing edges as well as the similarity of their subclass hierarchy. For case similarity, we compute a measure of importance for each relation based on how common the relation is, allowing us to weight rarer relations or entities as being more important when selecting which cases to retrieve.
Identifying and Applying Prediction Paths: After retrieving cases of past events in the KG, EvCBR next begins to identify paths through the KG which can be used to make predictions. Given that the retrieved cases have cause and efect event pairs, we aim to discover prediction paths of relations in the KG that can be followed from the cause event entity to the efect’s properties for each case. The confidence of each of these prediction paths is then scored, with higher confidence being assigned to paths that lead to the correct entity more often across all the retrieved cases.
We next apply these learned prediction paths to make predictions about our new query event. From the new input event, we follow paths to make predictions about the unseen efect event’s properties. The confidence of each path, calculated based on the precision of following the paths in past cause-efect event pairs, is used to compute an aggregated score of predictions for each property of the efect event.
Using this path-based prediction method allows EvCBR to inherently provide more explainability than many embedding-based models. Additionally, our method of identifying and scoring prediction paths is performed on-the-fly rather than requiring any pretraining process.
Refining Predictions: Lastly, we introduce an additional step to refine our prediction results. The intuition of our refinement method is that if EvCBR’s predictions about the efect event are correct, we should be able to follow the procedure in the reverse direction to accurately “predict” the input cause event’s properties starting from the efect event. Performing this process in reverse gives us a way to ground the confidence of our predictions in terms of the “ground-truth” properties of the input cause event.
Evaluation: We evaluated the performance of EvCBR against baseline KG completion methods, encompassing a number of notable models using methods such as distance based embeddings, graph neural networks, rule-learning, and case-based reasoning. We find EvCBR to have superior performance for event prediction as well as competitive performance for more general 2-hop link prediction. In addition to showing strong performance in our evaluations, the fact that EvCBR requires no training makes it well suited to our event prediction use case, and more generally when applied to a KG like Wikidata which is frequently updated.
We plan to demonstrate our event prediction system using a Jupyter Notebook environment
along with visualizations of the graph and predictions. Our target audiences for this
demonstration are researchers and developers of decision support solutions, e.g., in finance [