Event sequence prediction method based on point process, graph calculation and large model fusion
By fusing time domain and spatial domain features in event sequence prediction, combining large language models and knowledge graphs, graph neural networks are used for node predictions, and the prediction results of point processes and graph neural networks are sorted and compared to solve existing problems. The technology solves the problems of insufficient dimensionality, event non-independence and insufficient event semantic information to achieve more accurate event sequence prediction.
Patent Information
- Application Number
- CN202411426230.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-14
- Publication Date
- 2025-01-21
AI Technical Summary
The prior art has problems such as insufficient dimensions, event non-independence and event semantic information in event sequence prediction, resulting in inaccurate prediction results.
By fusing the characteristics of the time domain and the spatial domain, combining the large language model and the knowledge graph, the graph neural network is used for node prediction, and the prediction results of the point process and the graph neural network are sorted and compared to generate more accurate event sequence predictions. .
It realizes the learning of multi-dimensional event process information, improves the accuracy and reliability of event prediction, and can better fit the occurrence of events in reality.
Smart Images
Figure CN119338003A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of machine learning, and in particular relates to an event sequence prediction method based on point process and graph calculation and large model fusion. Background Art
[0002] Time series analysis is one of the research directions in the field of machine learning, and event series prediction is a subfield of time series analysis. Different from general time series data with constant sampling time intervals, event series is the sequence data obtained by sampling discrete events in continuous time. It is usually non-uniformly distributed in the time domain and space domain, that is, the intervals between events are usually non-constant and have a certain degree of randomness. Each sampling point in the event sequence represents a naturally occurring event, recording the time and type of the event.
[0003] Event sequence prediction is classified based on the prediction target. Common time series analysis tasks are mainly divided into four categories, namely time series prediction, time series classification, time series clustering and anomaly detection. The task that this invention mainly focuses on belongs to time series prediction, that is, to predict what type of event will occur in the future based on the observed historical event sequence. Modeling the sequence is the main method to achieve the time series prediction task. Early research methods include point process algorithms based on traditional machine learning, such as Poisson process models and self-excitation process models. With the great achievements of deep learning in image processing and natural language processing, point processes have gradually evolved into point process algorithms based on deep learning. In deep learning, recurrent neural networks have good effects in solving text sequence problems in natural language processing, so point process algorithms based on recurrent neural networks have been developed. Due to the limited memory capacity of recurrent neural networks and sensitivity to input sequences, as the attention mechanism is proposed, point process algorithms based on attention mechanisms have been developed. The premise of the above research on event sequences is that events are independent of each other. Some studies have adopted an event prediction model based on dynamic relationship modeling.
[0004] With the development of large language models, some studies have used large language models to study event sequence prediction. The research paper is: Language Model Can Improve Event Prediction by Few-shot Abductive Reasoning. This method architecture first generates candidate prediction events through the event sequence model, then uses the large language model to complete abductive reasoning for the prompt words of the candidate events, uses the generated events to match historical events, retrieves similar events, and finally merges and sorts the predicted events and retrieved events through the sorting model, and selects the optimal event as the prediction result.
[0005] The point-by-point process method of traditional machine learning and the point-by-point process method of deep learning are both applied to the time domain for event sequences. However, the components of an event include time, place, and people. Therefore, the study of event sequences is only applied to the time domain, and the prediction of event sequences is somewhat one-sided. Secondly, event sequence data also contains a large amount of semantic information, and the above methods do not fully consider speech information.
[0006] The event sequence prediction method based on the large language model takes into account the semantic information in the event sequence data and the temporal information in the event sequence, but does not take into account the spatial domain information in the event sequence data. In addition, when constructing the event sequence model, the method assumes that the events are independent and does not take into account the correlation between events in reality. Summary of the invention
[0007] In order to overcome the shortcomings of the prior art, the present invention provides an event sequence prediction method based on the fusion of point process and graph calculation and a large model. By fusing the features of the time domain and the space domain, the learning of the model's multi-dimensional event process information is realized. The learning of the semantic information of the model events and the relationship information between the events is realized through a large language model and a knowledge graph. The fusion of point process information and graph neural network information is realized through a large language model, thereby more accurately predicting event sequences.
[0008] The technical solution adopted by the present invention to solve the technical problem is as follows:
[0009] Step 1: For event sequence data, encode the event sequence in the time domain to obtain the time domain coding features;
[0010] Step 2: For event sequence data, encode the event sequence in the spatial domain to obtain spatial domain encoding features;
[0011] Step 3: Fuse the temporal domain coding features obtained in step 1 with the spatial domain coding features obtained in step 2;
[0012] Step 4: Use the deep point process method to pre-train the model using the feature data fused in step 3, and obtain the event prediction set for the unknown event sequence through the pre-trained model;
[0013] Step 5: Construct a historical event map based on the time, place, and people of historical events;
[0014] Step 6: Generate a promote that can be understood by the large language model from the elements in the event graph constructed in step 5, and adjust the large language model to obtain a large language model of historical events;
[0015] Step 7: Use the large language model obtained in step 6 to encode the elements in the historical event graph constructed in step 5 to integrate the language information;
[0016] Step 8: Use the graph neural network to perform node prediction tasks on the historical event graph constructed in step 5 to achieve training and reasoning functions;
[0017] Step 9: After the reasoning process in step 8, a node prediction set is obtained. A promote that can be understood by the large language model is generated through the node prediction set. The cause reasoning is performed through the large language model constructed in step 6 to obtain the cause event based on the graph neural network prediction set.
[0018] Step 10: Generate a promote that can be understood by the large language model through the event prediction set obtained in step 4, and perform cause reasoning through the large language model constructed in step 6 to obtain the cause event based on the point process prediction set;
[0019] Step 11: Sort and compare the cause events of the graph neural network prediction set obtained in step 9 with the cause events of the point process prediction set obtained in step 10. The comparison method is to compare the data of the two event sets with historical events respectively, score and sort them according to the similarity, and take the result with the highest ranking as the predicted event.
[0020] Preferably, the encoding method adopts one-hot encoding, random encoding and the like.
[0021] Preferably, the fusion method in step 3 is a summation method.
[0022] Preferably, the large language model is an open source large language model.
[0023] Preferably, the large language model is GLM3 or Llama3.
[0024] Preferably, the graph neural network is GNN or GAT.
[0025] The beneficial effects of the present invention are as follows:
[0026] 1. The most easily conceivable and feasible approach to event sequence prediction is the point process method based on traditional machine learning and the point process method based on deep learning to predict event sequences in the time domain. However, in real life, the factors that affect the occurrence of events are not only in the time domain. The present invention uses feature fusion technology to fuse the time domain and space domain features, thereby realizing multi-dimensional event prediction, making the model closer to reality, and improving the accuracy of prediction.
[0027] 2. Event sequence data not only contains sequence information, but also contains semantic information. The popularization of large language models can effectively obtain semantic information in event sequence data. Existing event prediction based on large language models takes into account the semantic information and sequence information of independent events. Events in life are often not chain-like independent events, but graph data with associations. This invention targets the characteristics of event graphs, combines knowledge graphs, large language models, graph neural networks and other algorithms, and simulates the occurrence of events in reality in a more realistic way.
[0028] 3. Finally, in order to better improve the accuracy of prediction, the present invention integrates the relevant ideas of ensemble learning, compares and sorts the prediction results of the point process with the results of the graph neural network by similarity, and takes the best sorted result as the prediction result. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0030] The present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0031] The main purpose of this invention is to consider the dimensions of event occurrence from both the time domain and the space domain. At the same time, it is necessary to integrate the semantic information and sequence information in the event sequence to ensure that the event sequence prediction can repeatedly fit the real events under higher dimensional information, thereby improving the accuracy of event prediction.
[0032] The technical problem to be solved by the present invention is the multi-dimensional event sequence information fusion problem. The present invention needs to consider the sequence information of the event sequence in the time domain and the spatial domain, as well as the semantic information of the event itself, so as to improve the accuracy and reliability of event prediction. This technology integrates technologies such as knowledge graph, graph neural network, and large language model, and belongs to the cross-task of natural language processing and graph computing in the field of artificial intelligence.
[0033] like Figure 1 As shown, an event sequence prediction method based on point process, graph calculation and large model fusion includes the following steps:
[0034] Step 101: Encode the event sequence data in the time domain. The encoding method may be one-hot encoding, random encoding, etc.;
[0035] Step 102: Encode the event sequence data in the spatial domain. The encoding method may be one-hot encoding, random encoding, etc.;
[0036] Step 103: Fusing the time domain coding features obtained in step 101 with the space domain coding features obtained in step 102, the fusion method is to fuse them in a summation manner;
[0037] Step 104: pre-training the model using the deep point process method with the feature data fused in step 103, and obtaining an event prediction set for an unknown event sequence through the pre-training model;
[0038] Step 105: construct an event map of historical events by using the time, place, and people of the events;
[0039] Step 106: Generate a promote that can be understood by a large language model from the elements in the event graph constructed in step 105, and fine-tune the large language model to obtain a large language model about historical events. The large language model can be an open source large language model, such as GLM3, Llama3, etc.
[0040] Step 107: Encode the elements in the historical event graph constructed in step 105 using the event language model trained in step 106 to integrate language information;
[0041] Step 108: Use a graph neural network to perform node prediction tasks on the historical event graph constructed in step 105 to achieve training and reasoning functions, where the graph neural network can select a graph neural network such as GNN and GAT;
[0042] Step 109: After the reasoning process of step 108, a node prediction set is obtained, and a promote that can be understood by the large language model is generated through the node prediction set. The cause reasoning is performed through the event large language model constructed in step 106 to obtain the cause event based on the graph neural network prediction set;
[0043] Step 110: Generate a promote that can be understood by the large language model through the point process event prediction set obtained in step 104, and perform cause reasoning through the event large language model constructed in step 106 to obtain a cause event based on the point process prediction set;
[0044] Step 111: Sort and compare the cause events of the graph neural network prediction set obtained in step 109 with the cause events of the point process prediction set obtained in step 110. The comparison method is to compare the data of the two event sets with historical events respectively, score and sort them according to the similarity, and take the result with the highest ranking as the predicted event.
[0045] The key point of this invention is the integration of multiple technologies, including point process, knowledge graph, large language model, graph neural network and other technologies. Through the integration of multiple technologies, the problems of insufficient dimension, non-independence of events and insufficient semantic information of events in event sequence prediction can be effectively solved. This makes the model simulate events in the real world more fully, thereby improving the accuracy of prediction.
[0046] The invention is intended to protect:
[0047] 1. Point process, graph computing engine, and event sequence prediction process integrated with large language model;
[0048] 2. Point process event prediction process integrating time domain and space domain features.
[0049] Commonly used event prediction methods mostly use point processes based on traditional learning and point processes based on deep learning. The core idea of the point process-based method is that events change over time, and event sequences are sequences with the time axis as the core. However, in reality, the factors that affect the occurrence of events are not only time, but also spatial information. Therefore, events should be seen as sequences in the time and space planes. Therefore, it is necessary to fuse the features in the time domain and the features in the space domain, and then perform the point process, so that the event process in reality can be more effectively imitated. At the same time, the event sequence data also contains a large amount of semantic information, and events in real life are not independent, so the assumption that events are independent has a certain one-sidedness. Therefore, it is necessary to construct an event knowledge graph, and by fusing a large language model with the knowledge graph, the semantic information of events during prediction and the relationship information between events are realized. The present invention realizes the learning of multi-dimensional event process information of the model by fusing the features of the time domain and the space domain, realizes the learning of the semantic information of model events and the relationship information between events through a large language model and a knowledge graph, and realizes the fusion of point process information and graph neural network information through a large language model, thereby more accurately predicting event sequences.
Claims
1. An event sequence prediction method based on point process, graph calculation and large model fusion, characterized in that: The steps include: Step 1: For event sequence data, encode the event sequence in the time domain to obtain the time domain coding features; Step 2: For event sequence data, encode the event sequence in the spatial domain to obtain spatial domain encoding features; Step 3: Fuse the temporal domain coding features obtained in step 1 with the spatial domain coding features obtained in step 2; Step 4: Use the deep point process method to pre-train the model using the feature data fused in step 3, and obtain the event prediction set for the unknown event sequence through the pre-trained model; Step 5: Construct a historical event map based on the time, place, and people of historical events; Step 6: Generate a promote that can be understood by the large language model from the elements in the event graph constructed in step 5, and adjust the large language model to obtain a large language model of historical events; Step 7: Use the large language model obtained in step 6 to encode the elements in the historical event graph constructed in step 5 to integrate the language information; Step 8: Use the graph neural network to perform node prediction tasks on the historical event graph constructed in step 5 to achieve training and reasoning functions; Step 9: After the reasoning process in step 8, a node prediction set is obtained. A promote that can be understood by the large language model is generated through the node prediction set. The cause reasoning is performed through the large language model constructed in step 6 to obtain the cause event based on the graph neural network prediction set. Step 10: Generate a promote that can be understood by the large language model through the event prediction set obtained in step 4, and perform cause reasoning through the large language model constructed in step 6 to obtain the cause event based on the point process prediction set; Step 11: Sort and compare the cause events of the graph neural network prediction set obtained in step 9 with the cause events of the point process prediction set obtained in step 10. The comparison method is to compare the data of the two event sets with historical events respectively, score and sort them according to the similarity, and take the result with the highest ranking as the predicted event.
2. The event sequence prediction method based on point process, graph calculation and large model fusion according to claim 1 is characterized in that: The encoding method adopts one-hot encoding, random encoding and the like.
3. The event sequence prediction method based on point process, graph calculation and large model fusion according to claim 1 is characterized in that: The fusion method in step 3 is a summation method.
4. The event sequence prediction method based on point process, graph calculation and large model fusion according to claim 1 is characterized in that: The large language model is an open source large language model.
5. The event sequence prediction method based on point process, graph calculation and large model fusion according to claim 1 is characterized in that: The large language models are GLM3 and Llama3.
6. The event sequence prediction method based on point process, graph calculation and large model fusion according to claim 1 is characterized in that: The graph neural network is GNN or GAT.