Future Event Prediction Based on Temporal Knowledge Graph Embedding

Accurate prediction of future events brings great benefits and reduces losses for society in many domains,such as civil unrest,pandemics,and crimes.Knowledge graph is a general language for describing and modeling complex systems.Different types of events continually occur,which are often related to historical and concurrent events.In this paper,we formalize the future event prediction as a temporal knowledge graph reasoning problem.Most existing studies either conduct reasoning on static knowledge graphs or assume knowledges graphs of all timestamps are available during the training process.As a result,they cannot effectively reason over temporal knowledge graphs and predict events happening in the future.To address this problem,some recent works learn to infer future events based on historical eventbased temporal knowledge graphs.However,these methods do not comprehensively consider the latent patterns and influences behind historical events and concurrent events simultaneously.This paper proposes a new graph representation learning model,namely Recurrent Event Graph ATtention Network(RE-GAT),based on a novel historical and concurrent events attention-aware mechanism by modeling the event knowledge graph sequence recurrently.More specifically,our RE-GAT uses an attention-based historical events embedding module to encode past events,and employs an attention-based concurrent events embedding module to model the associations of events at the same timestamp.A translation-based decoder module and a learning objective are developed to optimize the embeddings of entities and relations.We evaluate our proposed method on four benchmark datasets.Extensive experimental results demonstrate the superiority of our RE-GAT model comparing to various base-lines,which proves that our method can more accurately predict what events are going to happen.

News event prediction by trigger evolution graph and event segment

Event prediction aims to predict the most possible following event given a chain of closely related context events.Previous methods based on event pairs or the entire event chain may ignore much structural and semantic information.Current datasets for event prediction,naturally,can be used for supervised learning.Event chains are either from document-level procedural action flow,or from news sequences under the same column.This paper leverages graph structure knowledge of event triggers and event segment information for event prediction with general news corpus,and adopts the standard multiple choice narrative cloze task evaluation.The topic model is utilized to extract event chains from the news corpus to deal with training data bottleneck.Based on trigger-guided structural relations in the event chains,we construct trigger evolution graph,and trigger representations are learned through graph convolutional neural network and the novel neighbor selection strategy.Then there are features of two levels for each event,namely,text level semantic feature and trigger level structural feature.We design the attention mechanism to learn the features of event segments derived in term of event major subjects,and integrate relevance between event segments and the candidate event.The most possible next event is picked by the relevance.Experimental results on the real-world news corpus verify the effectiveness of the proposed model.

On the “spring predictability barrier” for strong El Nio events as derived from an intermediate coupled model ensemble prediction system

Using predictions for the sea surface temperature anomaly(SSTA) generated by an intermediate coupled model(ICM)ensemble prediction system(EPS), we first explore the "spring predictability barrier"(SPB) problem for the 2015/16 strong El Nio event from the perspective of error growth. By analyzing the growth tendency of the prediction errors for ensemble forecast members, we conclude that the prediction errors for the 2015/16 El Nio event tended to show a distinct season-dependent evolution, with prominent growth in spring and/or the beginning of the summer. This finding indicates that the predictions for the 2015/16 El Nio occurred a significant SPB phenomenon. We show that the SPB occurred in the 2015/16 El Nio predictions did not arise because of the uncertainties in the initial conditions but because of model errors. As such, the mean of ensemble forecast members filtered the effect of model errors and weakened the effect of the SPB, ultimately reducing the prediction errors for the 2015/16 El Nio event. By investigating the model errors represented by the tendency errors for the SSTA component,we demonstrate the prominent features of the tendency errors that often cause an SPB for the 2015/16 El Nio event and explain why the 2015/16 El Nio was under-predicted by the ICM EPS. Moreover, we reveal the typical feature of the tendency errors that cause not only a significant SPB but also an aggressively large prediction error. The feature is that the tendency errors present a zonal dipolar pattern with the west poles of positive anomalies in the equatorial western Pacific and the east poles of negative anomalies in the equatorial eastern Pacific. This tendency error bears great similarities with that of the most sensitive nonlinear forcing singular vector(NFSV)-tendency errors reported by Duan et al. and demonstrates the existence of an NFSV tendency error in realistic predictions. For other strong El Nio events, such as those that occurred in 1982/83 and 1997/98, we obtain the tendency errors of the NFSV structure, which cause a significant SPB and yield a much larger prediction error. These results suggest that the forecast skill of the ICM EPS for strong El Nio events could be greatly enhanced by using the NFSV-like tendency error to correct the model.

On Prediction of Record-Breaking Daily Temperature Events

The daily maximum/minimum temperature data at 740 stations in China from 1960 to 2005 were ana-lyzed to reveal the statistical characteristics of record-breaking（RB）daily extreme temperature events in the past 46 yr.It is verified that the observational daily extreme temperatures obey the Gaussian distribution. The expected values of RB extreme temperatures were obtained based on both the Gaussian distribution model and the initial condition of observed historical RB high/low temperature events after tedious the-oretical derivation.The results were then compared with those obtained by the iteration computation of the pure theoretical model.The comparison suggests that the results from the former are more consistent with the observations than those from the latter.Based on the above analyses,prediction of future possible RB high/low temperature events is made,and the spatial distributions of maximum/minimum theoretical values of their intensities are also given.It is indicated that the change amplitudes of future extreme temperatures differ evidently from place to place,showing a remarkable regional feature：the future extremely high temperature events will have a strong rising intensity in Southwest China,and a relatively weak rising intensity in western China;while the largest decrease of the future extremely low temperature events will appear in Northeast China and the north of Northwest China,and the decrease will be maintained relatively stable in space in Central China and Southwest China,in comparison with the historical low temperature pattern.Features in the occurrence time of the future RB temperature events are also illustrated.