A self-explainable graph machine learning approach for node classification

Abstract

A computer-implemented method provides self-explainable machine learning node classification for a machine learning model that encodes a graph in a latent space. The graph has nodes in an original input space and edges that semantically connect pairs of nodes. The method includes embedding learned prototype nodes into the latent space of the model, using the model to predict a class for a particular node, determining a similarity score between the latent representation of the particular node in the latent space and each of the learned prototype nodes, mapping the node's latent representation and the learned prototype nodes back to the original input space, and providing a predicted class explanation for the particular node based on the mapped prototype nodes. The present invention can be used in a variety of applications, including, but not limited to, several anticipated use cases in drug discovery, medical diagnostics / applications, and medicine.

Description

[Technical Field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This patent application claims priority to European Patent Application No. EP 23154902.3, filed on February 03, 2023, the entire disclosure of which is incorporated herein by reference.

[0002] The present disclosure relates to methods, systems, and computer-readable media for building prototypes of self-explainable graph machine learning techniques for node classification. [Background technology]

[0003] Artificial intelligence (AI) has emerged as a transformative technology with an ever-expanding role in making accurate predictions across a variety of domains. AI's usefulness lies in its unparalleled ability to process massive amounts of data, uncover hidden patterns, and extrapolate insights, enabling business and industrial managers to make informed decisions. From weather forecasting to medical diagnosis, AI has proven to be a powerful tool for increasing the accuracy and reliability of predictions. The integration of machine learning algorithms allows AI systems to continuously improve their predictive capabilities as they learn from new data, allowing them to adaptively and efficiently keep up with dynamic environments. As a result, AI-powered predictive models have become indispensable assets, fueling innovation, optimizing processes, and enabling individuals and organizations to proactively respond to future scenarios.

[0004] However, there are still problems with assessing the reliability of AI system predictions. To this end, an emerging requirement for creating trustworthy AI systems is the ability to provide explainable predictions. Here, the concept is to provide explanations to humans to gain confidence in the predictions of machine learning models, and the rationale must be understandable to humans. However, human understandability is difficult to achieve, especially when the input data is structured as a graph of interconnected entities and each individual entity needs to be classified.

[0005] Furthermore, in data-sensitive domains like healthcare, an additional requirement must be met: the explanation cannot include sensitive, patient-related data. This is a very challenging requirement for current graph AI methods, which typically provide explanations of node classifications in terms of other nodes and connections in the graph. In principle, this could raise serious privacy issues, including violating the European Union's General Data Protection Regulation (GDPR). Therefore, finding alternatives that comply with privacy laws (e.g., comply with the GDPR) is crucial. [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] Zhang et al., “ProtGNN: Towards Self-Explaining Graph Neural Networks”, AAAI-22, 9127~9135 (2022) [Non-patent document 2] Shin et al., “PAGE: Prototype-Based Model-Level Explanations for Graph Neural Networks”, arXiv, 1~18 (2022) [Non-patent document 3] Dai and Wang, "Towards Prototype-Based Self-Explainable Graph Neural Network," Assn. for Computing Machinery Conference 2017, 1–9 (2017). [Non-patent document 4] Magister et al., "Encoding Concepts in Graph Neural Networks," arXiv, 1–24 (August 2022) [Non-Patent Document 5] Huang et al., "GraphLIME: Local Interpretable Model Explanations for Graph Neural Networks," IEEE Transactions on Knowledge and Data Engineering, 35:7, 6968–6972, (June 2022) [Non-patent document 6] Ying et al., “GNNExplainer: Generating Explanations for Graph Neural Networks,” 33rd Conference on Neural Information Processing Systems, 1–13 (2019) [Non-Patent Document 7] Jiang et al., “Variational Deep Embedding: An Unsupervised and Generative Approach to Clustering,” IJCAI 2017, 1–22 (2017) Summary of the Invention [Means for solving the problem]

[0007] According to one aspect of the present disclosure, a computer-implemented method for self-explainable machine learning node classification is provided for a machine learning model that encodes a graph in a latent space. The graph has a set of nodes in an original input space and a set of edges semantically connecting pairs of nodes. The method includes embedding learned prototype nodes into the latent space of the model, using the model to predict a class from a set of classes for at least certain of the nodes, determining a similarity score between a latent representation of the certain node in the latent space and each of the learned prototype nodes, mapping the node latent representations and the learned prototype nodes to the original input space, and providing a predicted class explanation for at least the certain node based on the mapped prototype nodes in the original input space. The method has applications including, but not limited to, AI-assisted drug development (AIDD) or digital medicine and personalized medicine for patient prediction, public safety, smart cities, and other machine learning applications that use graphs.

[0008] The present invention may be used in a variety of applications, including but not limited to, drug discovery, medical diagnostics / applications, and in medicine, some anticipated use cases.

[0009] The embodiments of the present disclosure will be described in further detail below based on the exemplary drawings. The present disclosure is not limited to the exemplary embodiments. All features described and / or illustrated herein can be used alone or in different combinations in the embodiments of the present disclosure. The features and advantages of various embodiments of the present disclosure will become apparent by reading the following detailed description with reference to the accompanying drawings, in which: [Brief explanation of the drawings]

[0010] [Figure 1] 1 illustrates an exemplary embodiment of a system and method implemented in accordance with aspects of the present disclosure. [Figure 2] FIG. 1 is a block diagram of an example processing system that may be configured to perform any of the operations disclosed herein. DETAILED DESCRIPTION OF THE INVENTION

[0011] Aspects of the present disclosure provide a self-explainable graph machine learning technique for node classification tasks that can also generate human-understandable predictions via prototypes associated with target classes. The disclosed machine learning technique iteratively learns from human input, updates its prototypes, and updates its explanations accordingly.

[0012] Compared to the current state of the art, the method provided in this disclosure assumes that humans are poor at interpreting subgraphs as explanations for node prediction tasks and instead focuses on generating new, connected nodes that represent "ideal candidates" for specific output classes. Additionally, aspects of the present disclosure provide a simple mechanism for incorporating human feedback into the loop and improving explanations, which cannot be employed in explainable graph artificial intelligence (AI) approaches that do not employ an autoencoder architecture. By simultaneously predicting classes and decoding prototype representations in the original graph space, embodiments implemented in accordance with the present disclosure can increase a user's confidence level. Such users can also seek additional explanations by generating variants of the prototype representation from associated distributions; in contrast, literature methods generally generate a fixed amount of possible explanations.

[0013] Embodiments implemented in accordance with the present disclosure provide technical improvements in generating human-understandable predictions in node classification scenarios. For example, embodiments implemented in accordance with the present disclosure provide at least the following capabilities ("desiderata") that are not generally available in existing systems:

[0014] Easy to understand explanation: Most existing methods simply return a subgraph that is deemed useful for predicting node / graph classes. For example, Zhang et al., "ProtGNN: Towards Self-Explaining Graph Neural Networks," AAAI-22, 9127-9135 (2022) (<<https: / / ojs.aaai.org / index.php / AAAI / article / download / 20898 / 20657> >), Shin et al., "PAGE: Prototype-Based Model-Level Explanations for Graph Neural Networks", arXiv, 1-18(2022)(available at << https: / / arxiv.org / pdf / 2210.17159.pdf >>), Dai and Wang, "Towards Prototype-Based Self-Explainable Graph Neural Network", Assn. for Computing Machinery Conference 2017, 1-9(2017)(<<https: / / arxiv.org / pdf / 2210.01974.pdf> Magister et al., "Encoding Concepts in Graph Neural Networks," arXiv, 1–24 (August 2022) (available at <<https: / / arxiv.org / pdf / 2207.13586.pdf> >, the entire contents of each of which are incorporated herein by reference. In a typical situation, a human without precise domain knowledge cannot understand the meaning of the subgraph and is therefore confused. In contrast, embodiments of the present disclosure provide a human-understandable explanation of the graph.

[0015] Model-level explanations: Most existing methods rely on post-hoc explanations, i.e., explanations are derived after the model has been trained. For example, Huang et al., "GraphLIME: Local Interpretable Model Explanations for Graph Neural Networks," IEEE Transactions on Knowledge and Data Engineering, 35:7, pp. 6968-6972, (June 2022) (available at: https: / / ieeexplore.ieee.org / stamp / stamp.jsp?tp=&arnumber=9811416); Ying et al., "GNNExplainer: Generating Explanations for Graph Neural Networks," 33rd Conference on Neural Information Processing Systems, pp. 1-13 (2019) (available at:<https: / / arxiv.org / pdf / 1903.03894.pdf> (Available at: http: / / www.ncbi.nlm.nih.gov / pubmed / 1024777) (the entire contents of each of which are incorporated herein by reference). In contrast, embodiments of the present disclosure learn to generate explanations along with decisions during training.

[0016] Generating the required number of explanations: Existing methods exist that provide a set of explanations for a given prediction, but typically do not provide additional explanations. On the one hand, generating too many explanations can confuse humans, and on the other hand, providing too few explanations may not be enough for humans to trust the prediction system. Embodiments of the present disclosure, however, can address scenarios where humans may not immediately understand a given explanation by suggesting new, but still plausible, alternative reasons why a prediction was made.

[0017] Embodiments implemented according to the present disclosure are enabled to provide at least the aforementioned functionality, at least in part, by implementing one or more of the following technical improvements:

[0018] Providing explanations as learned prototypical nodes in the graph: Embodiments of the present disclosure learn "template / prototypical nodes" that represent specific classes. They present humans with a simple interpretation in the original space of features: if the node features are easily understandable to humans, then the prototypical nodes are also easily understandable to humans.

[0019] The prototype nodes remain in the embedding space: embodiments of the present disclosure train the model to generate these prototypes while also classifying the nodes. The prototypes are embedded in the latent space of the model and can then be decoded as newly generated nodes in the original graph. In this way, post-hoc explanations are avoided and explanations can function to predict specific classes.

[0020] Prototype nodes are represented as parameters of a distribution: In an embodiment of the present disclosure, prototype nodes are encoded in the latent space as parameters of some distribution, e.g., a Gaussian. In this way, by sampling from this particular Gaussian, it is possible to consider the most likely prototype node or generate less "probable" explanations. Thus, there is no limit to the number of explanations a user can seek.

[0021] The set of three properties above is not present in existing systems. An embodiment of the present disclosure implements these properties with a variational graph auto-encoder (VGAE) in a latent variable modeling framework. A VGAE is a machine learning model trained to find low-dimensional, yet meaningful, representations of nodes. Unlike previous work, a VGAE is constructed from a generative distribution, defined as a mixture of K distributions. The variational distribution attempts to approximate the true posterior of the data and is instead implemented in a more classical way.

[0022] Embodiments of the present disclosure also utilize unstructured (i.e., vector) input data that relies on the idea of ​​prototypes, or representations of prototypes, which are abstract concepts that cannot be identified with specific entities (e.g., patients) while conveying useful information.

[0023] According to a first aspect of the present disclosure, a method for implementing a self-explainable graph machine learning scheme for node classification is provided. The method learns prototype representations in a latent space to interpret predictions of graph entities by a variational graph autoencoder. The prototype representations can be mapped from the latent space to the original graph space by a trained decoder for easy inspection by a user. The method can generate a potentially infinite number of prototype representations to increase a user's confidence level by sampling from distributions associated with the latent prototypes. The method can provide human-understandable evidence that the model has learned to associate a particular prototype representation with a given output class by examining the weights of an interpretable classifier built on the prototypes and the similarity of the input with respect to the prototypes.

[0024] According to a second aspect of the present disclosure, there is provided a method for prototype-based, self-explainable node classification, the method comprising: 1) recording or collecting information (e.g., from a patient network) and pre-processing the information into a graph with node features; 2) Embedding entities into the latent space learned by the method; 3) making interpretable predictions for individual entities by comparing them with learned prototype representations; 4) Providing an explanation for the prediction in terms of a representation of the decoded prototype as connected nodes in the original graph; Includes.

[0025] In one embodiment of one of the aspects, the method may further include iteratively updating the explanation based on input relating to any domain.

[0026] According to a third aspect of the present disclosure, a computer-implemented method for self-explainable machine-learned node classification is provided for a machine-learning model that encodes a graph in a latent space. The graph has a set of nodes in an original input space and a set of edges semantically connecting pairs of nodes. The method includes embedding learned prototype nodes into the latent space of the model, using the model to predict classes from a set of classes for at least certain of the nodes, determining a similarity score between a latent representation of the certain nodes in the latent space and each of the learned prototype nodes, mapping the node latent representations and the learned prototype nodes to the original input space, and providing a predicted class explanation for at least the certain nodes based on the mapped prototype nodes in the original input space.

[0027] In an embodiment of the third aspect, the method may further include recording, collecting, or obtaining information from the target area and processing the information into a graph in the original input space.

[0028] In one embodiment of the third aspect, embedding the learned prototype nodes in the latent space of the model includes generating K distributions in the latent space and determining K means, respectively, from one of the K distributions, where each of the K means is one of the learned prototype nodes. The method may further include encoding the graph in the latent space, where for each respective node of the graph in the original input space, generating a respective distribution in the latent space, where the mean of each distribution is a respective latent representation of the respective node, and determining a probability that each latent representation of the node is generated by one of the K distributions.

[0029] In one embodiment of the third aspect, K may be a predetermined integer given by a user, and each of the K distributions is a continuous distribution.

[0030] In one embodiment of the third aspect, the encoder performs graph encoding in a latent space, and the encoder has a graph convolutional network or a graph attention network.

[0031] In one embodiment of the third aspect, predicting the class of a particular node comprises obtaining a K-component similarity vector from the similarity scores, and processing the similarity vector with a linear classifier to predict a class for the particular node.

[0032] In one embodiment of the third aspect, the method further comprises examining the weights of the linear classifier, the relative weights giving a relative indication of how important similarity to one of the nodes of the learned prototype is for predicting the class.

[0033] In one embodiment of the third aspect, the linear classifier includes an objective function with a regularization term to select the sparse similarity vector.

[0034] In one embodiment of the third aspect, a trained decoder maps the latent representations of the nodes and the learned prototype nodes back to the original input space. The decoder may also include a trained graph machine learning model based on a loss that trains the decoder to reconstruct the original node features from the latent representations.

[0035] In an embodiment of the third aspect, the graph machine learning model is further configured to reconstruct the original edges between the nodes using the trained graph autoencoder.

[0036] In an embodiment of the third aspect, the method may further include providing an alternative explanation for the prediction of the class by selecting, from among the K distributions, a distribution determined to be associated with the class; sampling a vector of the latent space from the selected distribution; mapping the sampled vector to the original space; and providing an alternative explanation based on the sampled vector mapped to the original space.

[0037] In one embodiment of the third aspect, a decoder may be provided, comprising a trained machine learning model that performs a mapping of prototypical nodes to the original input space. The method may further include receiving input from a user suggesting a change to a feature of a selected one of the prototypical nodes mapped to the original input space, and retraining only the decoder based on the input such that the selected prototypical node is decoded as indicated by the user.

[0038] In one embodiment of the third aspect, the target domain is a patient network where the information of the target domain is partially available data of patients including vital measurements and clinical test results, the learned prototype nodes represent learned prototype patients associated with clinical diagnoses, and the set of classes is a set of diseases.

[0039] As will be appreciated by those skilled in the art, the embodiments of the third aspect may be combined in various ways, all within the scope of the present disclosure.

[0040] According to a fourth aspect of the present disclosure, there is provided a computer system including one or more hardware processors, configured alone or in combination, to enable execution of the methods of each of the above aspects, including variously implemented or combined embodiments.

[0041] According to a fifth aspect of the present disclosure, there is provided a tangible, non-transitory, computer-readable medium having instructions thereon that, when executed by one or more hardware processors, enable the execution of the methods of each of the above aspects, alone or in combination, including variously implemented or combined embodiments.

[0042] The above and further features and advantages of various embodiments of the present disclosure will be apparent from the following description of exemplary embodiments, which are given with reference to the drawings.

[0043] 1 illustrates one embodiment of a method and system 100 implemented in accordance with the present disclosure. To address at least the above-mentioned desiderata, the method implements a variational formulation of the problem and tackles three different machine learning tasks to be optimized using backpropagation: 1) node classification, 2) node reconstruction, and 3) link prediction.

[0044] Input and Problem Definition As shown in Figure 1, in a first step 101, a set of input data is collected or possibly obtained. The input includes a collection of entities that are semantically connected to each other so that they can be represented as a graph. For example, in a second step 102, the collected data may undergo pre-processing operations to, among other things, create a graph of the collected data.

[0045] A graph g (sometimes called a network) is a machine-understandable data structure that provides a semantically structured representation of information. A graph represents entities (e.g., patients, doctors, symptoms, illnesses, vital signs, etc.) as "nodes" and relationships between entities (e.g., has symptoms, is being treated, has blood results, has weight, etc.) as "edges" in the graph. Essentially, two nodes connected by a relationship form the "thing" captured by the graph. A graph g may be characterized as a tuple (V, E, X, A), where V is a set of nodes (e.g., patients), E is a set of directed edges connecting pairs of nodes, X is the domain of node features, and (possibly) A is the domain of edge features. As used herein, the domain of node features refers to the set of possible attributes, characteristics, or properties associated with individual nodes in the graph, and the domain of edge features refers to the set of possible attributes, characteristics, or properties that may be attributed to an edge.

[0046] From the graph g, a feature vector can be derived. A feature vector is a structured representation of attributes or characteristics associated with either a node or an edge. Thus, a feature vector encapsulates relevant information about a particular entity or edge of the graph g. The feature vector can serve as input data for machine learning algorithms, allowing the algorithm to analyze and make predictions based on the given attributes. In this disclosure, x u ∈ X denotes the feature vector of node u (e.g., including patient measurements such as heart rate, blood information, weight, etc.).

[0047] The task (or problem) of the embodiment of FIG. 1 is to classify each node of the graph while providing an explanation of those classifications in the form of prototypical nodes.

[0048] A "prototype node" refers to a representative or characteristic node in a graph. A prototype node embodies a typical attribute or characteristic that is common among a subset of nodes in the graph. A prototype node exists in the original space of the graph and also has a corresponding representation in the latent space, which is an n-dimensional Euclidean space onto which the nodes in the original space of the graph are mapped, where n is typically (but not necessarily) a small number.

[0049] Prototype node generation distribution Below we describe the (learned) generative distribution of the variational graph autoencoder (VGAE) method implemented in this disclosure.

[0050] The generating distribution of the VGAE method is defined as a mixture of K distributions in a latent space of dimension d. For example, if the user selects K=3 and Gaussian distributions, then the individual Gaussians

[0051]

number

[0052] The average of x, y ...

[0053] Encode the nodes in the latent space Returning to FIG. 1, in the third step 103, the encoder 110 implements the variational distribution according to the VGAE framework provided in this disclosure. That is, for each node u, the encoder 110 calculates the mean z u and generates a distribution whose nodes have a probability of belonging to one of the K components of the mixture of distributions. The encoder 110 may be implemented in various ways using graph machine learning techniques, such as with a graph convolutional network (GCN) and a graph attention network (GAT). However, the present disclosure is not limited to a particular machine learning model and may be implemented with a machine learning model configured for graph-structured data that transforms nodes of an input graph into points in a latent space.

[0054] Interpretable predictions per node To perform interpretable predictions about node u, the concept of "prototype matching" is used. Using probabilistic techniques (e.g., using Bayes' theorem), z u and the prototype representation μ 1 , μ 2 , μ 3 A similarity score is calculated between each of the K=3 distributions. More specifically, the probability that the node is "generated" by one of the K=3 distributions (e.g., Gaussian) is calculated. Thus, a K-component similarity vector is obtained, which is fed into a linear classifier (e.g., a logistic regressor) to assign class y for a particular node u. u As shown in Figure 1, a predictor 120 performs a fourth step 104 that performs the above functions to classify the nodes.

[0055] The insight of this disclosure is that the weights of a linear model can be examined to determine how important similarity to a prototype is for predicting a class. To further improve the comprehensibility of the prediction, a regularization term may be added to the objective function of the classifier that selects the sparse similarity vector. The objective function may be implemented, for example, using the variational lower bound function (ELBO), which is well understood in the art. The regularization term minimizes the entropy of the variational distribution that computes the cluster assignment probability for each node in the input graph.

[0056] A linear classifier is a prototypical input representation μ x If the score for a given class is high enough (e.g., meets a user-imposed threshold), the system determines that the prototype is "particularly important" to that particular class.

[0057] Decrypting the original graph In a fifth step 105, the trained decoder 130 calculates the u The latent representation of the original x u The decoder 130 may be implemented, for example, with a multilayer perceptron (MLP) or another graph machine learning model. The loss trains the decoder to reconstruct the original node features from the latent representation. In addition, the decoder 130 may include a link prediction decoder, such as a graph autoencoder (GAE), to reconstruct the original links between nodes. Here, a binary classification loss may be used in conjunction with negative sampling of non-existent links to train the GAE.

[0058] The decoder may be implemented using one of many well-understood machine learning models designed for vector data that transform latent representations of nodes in the input graph into node features in the original graph space. The machine learning model implemented for the decoder also reconstructs the connections between the nodes.

[0059] Explaining predictions through prototypes Also in the fifth step 105, the decoder 130 generates the prototype representation μ 1 , μ 2 , μ 3 to the original input space of the node, which allows the system to interpret the meaning of the representation of each learned prototype in a human-understandable form.

[0060] Additionally, by using a link prediction decoder for visualization purposes, it is possible to connect such new nodes (i.e., remapped node representations) to the most similar learned nodes (i.e., prototypical nodes). If a learned prototype is determined to be particularly important for a particular class (as determined in the fourth step 104 above), the user may be told that a "prototypical" representation for the particular class has been identified, and its feature values, e.g., "proto-patient," may be shown along with some properties (e.g., determined by the decoder) that are most likely to lead to its classification (e.g., cardiovascular problem). This, along with the similarity vector for the particular node, helps humans understand what the model has learned.

[0061] Additionally, the user may request a different explanation to represent a certain class c. If so, the system selects some of the distributions from the generative model that appear to be relevant to class c (e.g., those above a "particularly important" threshold), samples values / vectors in the latent space from one or all of them, and decodes them as nodes in the original graph space as was done before.

[0062] Improving the representation of prototypes through new inputs Upon examining the learned prototype representations, the system allows the user to provide input by suggesting modifications to some features of the prototype representation (e.g., the input may indicate that the value for a feature predicted by the decoder should be higher / lower in a particular prototype representation). In that case, the system obtains this input by retraining only the node feature decoder 130 while simultaneously ensuring that the latent representation of the particular prototype is decoded as expected. This is represented as the sixth step 106 in FIG. 1. Decoder modifications do not change the latent representation of the node, and therefore improve explanation quality without altering classification performance.

[0063]

[0006] Embodiments of the present invention thus provide a self-explainable graph machine learning approach, thereby providing an overall improvement to computer in machine learning systems. Furthermore, embodiments of the present invention may be applied to use cases that provide further improvements in several technology areas that use or could benefit from graph-based machine learning, including, but not limited to, healthcare (e.g., digital medicine, personalized medicine, AI-assisted drug or vaccine development (AIDD or oncoimmunity), etc.), materials development, public safety, and smart cities (e.g., automated traffic or vehicle control, smart districts, smart buildings, smart industrial plants, smart agriculture, energy management, etc.).

[0064] Exemplary Implementation 1: Explaining Clinical Risk Prediction in a Patient Network Use Case: Using AI to predict clinical risk in hospitals can speed up the diagnosis of patients prone to a determined disease (e.g., sepsis, acute renal failure, etc.). However, doctors may be reluctant to blindly trust AI predictions without some intuitive knowledge / explanation of how the prediction process works. One implementation of the system and method according to the present disclosure can be used to maintain the same prediction accuracy as other methods, while also providing an intuitive explanation (through a prototypical representation) of why a patient was classified that way. The prediction can be fed into a prediction system, which outputs a set of additional tests to perform, for example, by providing the configuration of a complete blood count (CBC) blood test machine, along with the clinical risk prediction. Additionally, doctors can provide feedback on the prototypical representation, which can be easily incorporated into the system.

[0065] Data Source: A patient network consisting of partially available data of a patient. This data includes, but is not limited to, basic vital measurements and laboratory tests such as heart rate, oxygen saturation, weight, height, glucose, temperature, pH, etc. Patient blood samples are available for future use.

[0066] Methodology: A system implemented according to the present disclosure predicts each patient's clinical risk through its association with learned prototype patients. This association is presented to the physician in the form of an explanation, e.g., the prototype patient is a good representative of patients with heart disease, and the current patient is "close" to it in the latent space. This explanation provides insight into why a decision was made by the system and increases the physician's level of confidence. Additionally, the physician can gain further insight by requesting more prototype representations to be generated from the primary ones, which are intended to further convince the physician that the proposed system has learned meaningful associations between patient characteristics and clinical risk.

[0067] Output: Information about one or more prototype patients representing their particular clinical risk as a form of clinical risk prediction and explanation.

[0068] Exemplary Implementation 2: Patient Stratification: Predicting Response to Treatment in Clinical Trials and Identifying Ideal Candidates Use Case: One of the factors that most influences clinical trial budgets and ultimately their outcomes is the patient selection process. Often, patients who are unsuitable for treatment (because the trial does not yield a positive response) are enrolled anyway, which can cause significant financial losses that ultimately lead to the failure of the trial itself. The rationale is to use information at the genetic level to stratify patients and predict their early response, which results in a stratification between responders and non-responders. Given a network of patients connected according to their cell type distribution and partial knowledge of their response to a specific drug, the goal is to predict the response for other patients as well as identify "ideal candidates." Here, the assumptions include that patients with similar distributions of cell types respond similarly to the drug and that the set of gene-cell information correlates with such distributions.

[0069] Data sources: patient networks with genomic activity information (e.g., single-cell RNA sequencing), patient responses to specific drugs, and publicly available ontologies that can be used to build the graph.

[0070] Methodology: A system implemented according to the present disclosure predicts a patient's response to a drug according to the patient's similarity to a learned "ideal candidate." The method generates ideal responders and non-responders and embeds them in a patient graph that can be used by a user (study administrator) to verify the results.

[0071] Output: Prediction of patient response to medication (low / medium / high).

[0072] 2 illustrates a processing system 200 that may represent each of the computing systems disclosed herein. The processing system 200 may include one or more processors 202, memory 204, one or more input-output devices 206, one or more sensors 208, one or more user interfaces 210, and one or more actuators 212.

[0073] Processor 202 may include one or more separate processors, each having one or more cores. Each of the separate processors may have the same or different architecture. Processor 202 may include one or more central processing units (CPUs), one or more graphics processing units (GPUs), circuitry (e.g., application specific integrated circuits (ASICs)), digital signal processors (DSPs), etc. Processors 202 may be mounted on a common substrate or on multiple different substrates.

[0074] Processor 202 is configured to perform (e.g., configure to provide for the performance of) a function, method, or operation when at least one of one or more of the separate processors is capable of performing the operations that embody the function, method, or operation. Processor 202 can perform the operations that embody a function, method, or operation, for example, by executing code stored in memory 204 (e.g., interpreting a script) and / or passing data through one or more ASICs. Processor 202, and therefore processing system 200, can be configured to automatically perform any of the functions, methods, and operations disclosed herein. Thus, processing system 200 can be configured to implement any (e.g., all) of the protocols, devices, mechanisms, systems, and methods described herein.

[0075] For example, when this disclosure states that a method or device performs task "X" (or that task "X" is performed), such statement should be understood to disclose that processing system 200 can be configured to perform task "X." Processing system 200 is configured to perform a function, method, or operation when at least processor 202 is configured to perform it.

[0076] The memory 204 may include volatile memory, non-volatile memory, and any other medium capable of storing data. The volatile memory, non-volatile memory, and any other type of memory may include multiple different memory devices in multiple separate locations, each with a different structure. The memory 204 may include remotely hosted (e.g., cloud) storage.

[0077] Examples of memory 204 include non-transitory computer-readable media such as RAM, ROM, flash memory, EEPROM, any type of optical storage disk such as a DVD, Blu-Ray® disk, magnetic storage, holographic storage, HDD, SSD, or any medium that can be used to store program code in the form of instructions or data structures. Any of the methods, functions, and operations described herein may be embodied entirely in the form of tangible and / or non-transitory machine-readable code (e.g., interpretable script) stored in memory 204.

[0078] The input-output device 206 may include any component for passing data, such as a port, an antenna (i.e., a transceiver), a printed conductive path, etc. The input-output device 206 may enable wired communication via USB, DisplayPort, HDMI, Ethernet, etc. The input-output device 206 may enable electronic, optical, magnetic, and holographic communication with suitable memory 204. The input-output device 206 may enable wireless communication via Wi-Fi, Bluetooth, cellular (e.g., LTE, CDMA, GSM, WiMax, NFC), GPS, etc. The input-output device 206 may include wired and / or wireless communication paths. The input device may be a user device such as a smartphone, tablet, laptop, or personal computer, among others.

[0079] The sensors 208 can capture physical measurements of the environment and report them to the processor 202, e.g., a camera. The user interface 210 can include a display, physical buttons, a speaker, a microphone, a keyboard, etc. The actuators 212 can enable the processor 202 to control mechanical forces.

[0080] Processing system 200 may be distributed. For example, some components of processing system 200 may reside on a remotely hosted network service (e.g., a cloud computing environment), while other components of processing system 200 may reside on a local computing system. Processing system 200 may have a modular design in which some modules include multiple features / functionality as shown in FIG. 2 . For example, an I / O module may include volatile memory and one or more processors. As another example, individual processor modules may include read-only memory and / or a local cache.

[0081] While the subject matter of the present disclosure has been illustrated and described in detail in the drawings and the foregoing description, such illustration and description should be considered as illustrative or exemplary, and not limiting. The present invention is defined by the claims, and any statements herein that characterize the invention should likewise be considered as illustrative or exemplary, and not limiting. It will be understood by those skilled in the art that changes and modifications may be made within the scope of the following claims, which may include any combination of features from the different embodiments described above.

[0082] The terms used in the claims should be interpreted to have the broadest reasonable interpretation consistent with the above description. For example, the use of the articles "a" or "the" when describing an element should not be interpreted as excluding a plurality of elements. Similarly, the term "or" should be interpreted as inclusive, such that it does not exclude "A and B," unless it is clear from the context or the preceding description that the statement "A or B" refers to only one of A and B. Furthermore, the statement "at least one of A, B, and C" should be interpreted as one or more of the group of elements consisting of A, B, and C, and should not be interpreted as requiring at least one of each of the listed elements A, B, and C, regardless of whether A, B, and C are categorically related or not. Furthermore, references to "A, B, and / or C" or "at least one of A, B, or C" should be interpreted to include only any one entity from the listed elements, e.g., A, any subset from the listed elements, e.g., A and B, or the entire list of elements A, B, and C.

[0083] Additional relevant background information regarding AI models can be found, for example, in Jiang et al., "Variational Deep Embedding: An Unsupervised and Generative Approach to Clustering," IJCAI 2017, 1-22 (2017) (the entire contents of which are incorporated herein by reference).<https: / / arxiv.org / pdf / 1611.05148.pdf> can be found at: [Explanation of symbols]

[0084] 100 systems 110 Encoder 120 Predictor 130 Decoder 200 Processing Systems 202 processors 204 memory 206 Input-Output Devices 208 Sensors 210 User Interface 212 Actuator

Claims

1. 1. A computer-implemented method for self-explainable machine-learned node classification for a machine-learning model that encodes a graph in a latent space, the graph comprising a set of nodes in an original input space and a set of edges semantically connecting pairs of the nodes, the method comprising: Embedding the nodes of the learned prototypes into the latent space of the model; predicting a class from a set of classes for at least certain of said nodes using said model; determining a similarity score between the latent representation of the particular node in the latent space and each of the learned prototype nodes; mapping the latent representations of the nodes and the learned prototype nodes back to the original input space; providing a predicted class description for at least the particular node based on the mapped prototype nodes of the original input space; A method comprising:

2. recording, collecting, or obtaining information from a target region and processing said information into said graph in said original input space; The method of claim 1 further comprising:

3. embedding the learned prototype nodes into the latent space of the model, generating K distributions in the latent space; determining each of K means from one of the K distributions, each of the K means being one of the nodes of the learned prototype; Including, The method further comprises encoding the graph in the latent space, the encoding of the graph comprising: For each respective node of the nodes of the graph in the original input space: generating respective distributions in the latent space, the mean of each distribution being a respective latent representation for each node; determining the probability that the respective latent representations of the nodes are generated by one of the K distributions; The method according to claim 1 or 2, comprising:

4. K is a predetermined integer provided by the user, each of the K distributions is a continuous distribution; The method of claim 3.

5. 4. The method of claim 3, wherein an encoder performs the encoding of the graph in a latent space, the encoder comprising a graph convolutional network or a graph attention network.

6. predicting the class of the particular node, obtaining a similarity vector of K components from the similarity scores; processing the similarity vector through a linear classifier to predict the class for the particular node; 6. The method of any one of claims 1 to 5, comprising:

7. examining the weights of the linear classifier, the relative weights giving a relative indication of how important similarity to one of the nodes of the learned prototype is for predicting the class.

7. The method of claim 6, further comprising:

8. The method of claim 6 or 7, wherein the linear classifier includes an objective function with a regularization term to select sparse similarity vectors.

9. 9. The method of claim 1, wherein a trained decoder performs the mapping of the latent representations of the nodes and the learned prototype nodes to the original input space, the decoder comprising a trained graph machine learning model based on a loss that trains the decoder to reconstruct the original node features from the latent representations.

10. 10. The method of claim 9, wherein the graph machine learning model is further configured to reconstruct original edges between the nodes using a trained graph autoencoder.

11. The method comprises: alternative explanations for the prediction of the class, selecting a distribution from among the K distributions that is determined to be associated with the class; Sampling vectors in the latent space from the selected distribution; mapping the sampled vectors back to the original space; providing the alternative explanation based on the sampled vectors mapped into the original space; and Steps provided by 11. The method of any one of claims 1 to 10, further comprising:

12. a decoder including a trained machine learning model performs the mapping of the prototype nodes to the original input space, and the method comprises: receiving input from a user suggesting a change to a feature of a selected one of the prototype nodes mapped to the original input space; retraining only the decoder based on the input such that the selected prototype nodes are decoded as indicated by the user; 12. The method of any one of claims 1 to 11, further comprising:

13. 13. The method of any one of claims 1 to 12, wherein the target domain is a patient network, the information of the target domain is partially available data of patients including vital measurements and clinical test results, the learned prototype nodes represent learned prototype patients associated with clinical diagnoses, and the set of classes is a set of diseases.

14. 1. A computer system including one or more hardware processors, the one or more hardware processors configured, alone or in combination, to enable execution of a method for self-explainable machine learning node classification for a machine learning model that encodes a graph in a latent space, the graph comprising a set of nodes in an original input space and a set of edges semantically connecting pairs of the nodes, the method comprising: Embedding the nodes of the learned prototypes into the latent space of the model; predicting a class from a set of classes for at least certain of said nodes using said model; determining a similarity score between the latent representation of the particular node in the latent space and each of the learned prototype nodes; mapping the latent representations of the nodes and the learned prototype nodes back to the original input space; providing a predicted class description for at least the particular node based on the mapped prototype nodes of the original input space; 2. A computer system comprising:

15. 1. A tangible, non-transitory computer-readable medium having instructions thereon that, when executed by one or more hardware processors, alone or in combination, provide for the execution of a method for self-explainable machine-learned node classification for a machine-learning model that encodes a graph in a latent space, the graph comprising a set of nodes in an original input space and a set of edges semantically connecting pairs of the nodes, the method comprising: Embedding the nodes of the learned prototypes into the latent space of the model; predicting a class from a set of classes for at least certain of said nodes using said model; determining a similarity score between the latent representation of the particular node in the latent space and each of the learned prototype nodes; mapping the latent representations of the nodes and the learned prototype nodes back to the original input space; providing a predicted class description for at least the particular node based on the mapped prototype nodes of the original input space; 1. A tangible, non-transitory computer-readable medium, including:

Images