Motivation. Knowledge Graphs (KGs) nowadays power many important applications
including Web search1, question answering [
Large-scale KGs are often automatically constructed and highly incomplete [
An important type of external knowledge that is common in practice and to the best
of our knowledge has not been explicitly considered so far is event log data. Events
naturally appear in many applications including social networks, smart cities, and
manufacturing. In social networks the nodes of a KG can be people and locations, and edges
can be friendship relations and places of birth [
In this work we define an event log for a KG as a set of sequences constituted of entities (possibly with repetitions) that may occur in the KG as nodes. Moreover, we assume that not every entity from a KG, but only what we call event entities can occur in logs. In the above: visited cities, traffic signals, and alarms are event entities. As we see later in the paper this separation of entities in a KG into event and non-event is important and practically motivated. We now illustrate an industrial KG and event log. 1 https://en.wikipedia.org/wiki/Knowledge_Graph ( HE ( HE ( LE
LE LE HE
LE LE LE
Jam
HE )
connectedTo ConveyorA
connectedTo hasSource hasSource
LE HE zero-shot type type entity
ConveyorB hasSource BoardJam ConveyorC hasSource type type
JamAlarm
EnergyEvent CoilJam
Fig. 1: Excerpt of a manufacturing KG and an event log.
Illustration of Scenarios. Consider an industrial KG that is inspired by a Siemens automated factory, that we will use later on for experiments, and that contains information about factory equipment, products, as well as materials and processes to produce the products. The KG was semi-automatically generated by parsing heterogeneous spreadsheets and other semi-structured data repositories and it is incomplete. In Figure 1 we depict a small excerpt from this KG where solid lines denote relations that are in the KG while dashed – the missing relations. The KG contains the topology of the conveyors A, B, and C and says that two of them (A and B) are connected to each other: hConveyorA; connectedT o; ConveyorBi. The KG also stores operator control specifications, in particular, event entities that the equipment can emit during operation. For example, CoilJ am, is an event entity and it can be emitted by conveyor C, i.e., hCoilJ am; hasSource; ConveyorC i. Event entities have further semantics described by the typing, e.g. CoilJ am is of type J amAlarm, severity levels, and possible emitting source locations. At the same time, the KG misses the facts that the conveyors A and C are connected in the factory; that BoardJ am is of type J amAlarm, and HighEnergy (HE) has the source ConveyorA and is of type EnergyEvent.
Additionally, in the example, we assume that an event log recorded during the operation of the factory consists of three following sequences over event entities: (HE; LE; LE; BoardJ am); (HE; LE; LE; CoilJ am); (LE; HE; LE; HE): Observe that the event log suggests that a jam typically occurs after a sequence of two consecutive low energy consumption (LE) events.
Problem Statement. An event log gives external knowledge to the KG by specifying frequent sequential patterns on the KG’s entities. These patterns capture some processes that the nodes of a KG can be involved in, i.e., manufacturing with machines described by the KG, traveling by a person mentioned in the KG, or traffic around traffic signals. This type of external knowledge has conceptual differences from text corpora where KG entities are typically described in a natural language and where occurrences of KG entities do not necessarily correspond to any process. Events are also different from rules or constraints that introduce formal restrictions on some relations.
The goal of this work is to understand how event logs can enhance relational learning for KGs. We address this problem by proposing an Event-enhanced Knowledge Learning (EKL) approach for KG completion that intuitively has two sub–steps: 1. Event alignment, where event entities are aligned in a low-dimensional vector space that reflects sequential similarity, and 2. KG completion, where the KG is extended with missing edges that can be either event-specific, e.g., such as the type edge between BoardJ am and J amAlarm in our rubbing example, or not event-specific, e.g., such as connectedTo between ConveyorA and ConveyorC in the running example.
Observe, the event logs directly influence the first step while also indirectly the second step of EKL. Hence, we expect a collective learning effect in a sense that the overall KG completion can benefit from event alignment, and vice versa.
Illustration of Ideas. During the first step EKL will align BoardJ am and CoilJ am to be similar. In the second step EKL will accordingly adjust entities ConveyorC and ConveyorB and then predict that ConveyorA is likely to also be connected to ConveyorC. Intuitively the missing link between the conveyors can be inferred from the sequential pattern in the event log: the log tells us that both BoardJ am and CoilJ am occur as a consequence of two consecutive LE events and therefore exhibit similar semantics. This similarity is carried to conveyor entities B and C, which leads to an increased likelihood that they both follow the same entity ConveyorA.
Note that the prediction of event-specific missing links is not the standard task for relational learning since we are predicting links within the background. Moreover, our approach can address the zero-shot scenario, where some event entities only appear in the event log, but they are novel to the KG (it is marked with red in Figure 1). E.g., HE in the running example corresponds to an entity that is missing in the KG, that has to be aligned during the first step of EKL and then linked to ConveyorA as well as to its type during the second step of EKL. Thus, EKL can also populate a KG with new (unseen) entities.
Contributions. The contributions of our work are as follows: – We proposed several EKL approaches to KG completion that comprise two model architectures that allow to combine (representations of) a KG and an event log for simultaneous training of both representations; this requires a nontrivial design of a model architecture that reflects interconnections of shared embeddings, three models for event logs that reflect different notions of event context. – We conducted an extensive evaluation of our approach and comparison to a stateof-the-art baseline on real-world data from a factory, on smart city traffic data, and controlled experiment data. Our results show that we significantly outperform two state-of-the-art baselines and the advantages are most visible for predicting links between entities that reflect the sequential process nature within the KG.
We presented a very preliminary version of this work as a short in-use paper [
Acknowledgements. This work was partially supported by the EPSRC projects DBOnto, MaSI3 and ED3.