Systems and methods to search for digital twins
To search for digital twins (avatars) on a computerized platform, a semantic embedding is made of the search query and this is compared with respective semantic embeddings generated for digital twins on the computerized platform, to determine a first set digital twins whose semantic embeddings are similar to that of the search query. A trained graph neural network generates respective graph embeddings of feature data of the first set of digital twins, and these graph embeddings are compared with graph embeddings that the trained graph neural network generates for other digital twins on the platform, to determine a second set of digital twins whose graph embeddings are similar to graph embeddings of the first set of digital twins. The search results returned in response to the search query may include at least one digital twin in the second set of digital twins.
1. A computer-implemented method of searching for digital twins in a target dataset on a computerized platform, said target dataset comprising a graph in which the digital twins are the nodes and the relations between them are the edges, the method comprising:
receiving a search query comprising one or more search terms;
generating a semantic embedding of the search query;
comparing the semantic embedding of the search query with respective semantic embeddings generated in respect of each of a plurality of digital twins on said computerized platform, to determine a first set of one or more digital twins having semantic embeddings similar to the semantic embedding of the search query;
inputting, into a trained graph neural network, feature data of the first set of one or more digital twins and data from said graph, to generate at least one respective graph embedding in respect of the first set of digital twins, said trained graph neural network comprising a graph neural network architecture trained using training data from said graph;
comparing the at least one graph embedding generated by the trained graph neural network in respect of the first set of digital twins with graph embeddings generated by said trained graph neural network in respect of other digital twins on said computerized platform, to determine a second set of one or more digital twins whose graph embeddings are similar to the at least one graph embedding of said first set of digital twins; and
outputting search results comprising information identifying at least one digital twin of said second set of digital twins.
2. The computer-implemented search method according to claim 1 , wherein the comparing of the semantic embedding of the search query with respective semantic embeddings generated in respect of each of a plurality of digital twins on said computerized platform comprises comparing of the semantic embedding of the search query with respective semantic embeddings of textual information characterizing the digital twins.
3. The computer-implemented search method according to claim 1 , wherein the semantic embedding of the search query and the semantic embeddings for said plurality of digital twins are generated using a Sentence-Bidirectional Encoder Representations from transformers (SBERT) network.
4. The computer-implemented search method according to claim 1 , wherein the graph neural network architecture is a GraphSAGE (SAmple and aggreGatE) neural network architecture.
5. The computer-implemented search method according to claim 4 , wherein the GraphSAGE neural network architecture is configured to process input feature sets comprising (i) data properties of digital twins, and (ii) semantic embeddings generated for said digital twins.
6. The computer-implemented search method according to claim 5 , wherein the semantic embeddings generated for the digital twins are semantic embeddings of textual information characterizing said digital twins, said semantic embeddings of textual information being generated by a Sentence-Bidirectional Encoder Representations from transformers (SBERT) network.
7. The computer-implemented search method according to claim 1 , comprising assigning a rank score to digital twins in said first and second sets of digital twins, and selecting or ordering the digital twins for output based on rank score.
8. The computer-implemented search method according to claim 1 , configured to search on an Internet-of-Things platform for digital twins corresponding to Internet-of-Things entities having respective assigned classes, the method further comprising:
detecting if plural digital twins included in said first and/or second sets of digital twins correspond to Internet-of-Things entities having the same assigned classes and, upon detecting such plural digital twins, discarding at least one of said plural digital twins from the search results.
9. A search/recommendation system configured to search for digital twins in a target dataset on a computerized platform, said target dataset comprising a graph in which the digital twins are the nodes and the relations between them are the edges, said system comprising:
a computing apparatus programmed to execute instructions to perform a method of searching for the digital twins in the target dataset, the method comprising:
receiving a search query comprising one or more search terms;
generating a semantic embedding of the search query;
comparing the semantic embedding of the search query with respective semantic embeddings generated in respect of each of a plurality of digital twins on said computerized platform, to determine a first set of one or more digital twins having semantic embeddings similar to the semantic embedding of the search query;
inputting, into a trained graph neural network, feature data of the first set of one or more digital twins and data from said graph, to generate at least one respective graph embedding in respect of the first set of digital twins, said trained graph neural network comprising a graph neural network architecture trained using training data from said graph;
comparing the at least one graph embedding generated by the trained graph neural network in respect of the first set of digital twins with graph embeddings generated by said trained graph neural network in respect of other digital twins on said computerized platform, to determine a second set of one or more digital twins whose graph embeddings are similar to the at least one graph embedding of said first set of digital twins; and
outputting search results comprising information identifying at least one digital twin of said second set of digital twins.
10. A non-transitory computer-readable medium comprising instructions which, when executed by a processor of a computing apparatus, cause the processor to perform a search method for searching for digital twins in a target dataset, said target dataset comprising a graph in which the digital twins are the nodes and the relations between them are the edges, the method comprising:
receiving a search query comprising one or more search terms;
generating a semantic embedding of the search query;
comparing the semantic embedding of the search query with respective semantic embeddings generated in respect of each of a plurality of digital twins on said computerized platform, to determine a first set of one or more digital twins having semantic embeddings similar to the semantic embedding of the search query;
inputting, into a trained graph neural network, feature data of the first set of one or more digital twins and data from said graph, to generate at least one respective graph embedding in respect of the first set of digital twins, said trained graph neural network comprising a graph neural network architecture trained using training data from said graph;
comparing the at least one graph embedding generated by the trained graph neural network in respect of the first set of digital twins with graph embeddings generated by said trained graph neural network in respect of other digital twins on said computerized platform, to determine a second set of one or more digital twins whose graph embeddings are similar to the at least one graph embedding of said first set of digital twins; and
outputting search results comprising information identifying at least one digital twin of said second set of digital twins.