Oracle International Corporation

Felix Schmidt

Nipun Agarwal

Craig Schelp

Matteo Casserini

Andrew Brownsword

Onur Kocberber

Stuart Wray

Arno Schneuwly

Saeid Allahdadian

Pravin Shinde

Milos Vasic

Benjamin Schlegel

Pit Fender

Guang-Tong Zhou

Manel Fernandez Gomez

Nikola Milojkovic

Renata Khasanova

Amin Suzani

Juan Fernandez Peinador

Rod Reddekopp

Hossein Hajimirsadeghi

Akhilesh Singhania

Sam Idicula

Boris Klots

Matthias Brantner

Nitin Kunal

Arun Raghavan

Tayler Hetherington

Yasha Pushak

Hamed Ahmadi

Vikas Aggarwal

Kantikiran Pasupuleti

John S Kowtko

Pavan Chandrashekar

Kenyu Kobayashi

Marija Nikolic

Pritam Dash

In an embodiment, a computer stores and operates a logic encoder that is an artificial neural network that infers a fixed-size encoded logic from textual or tokenized source logic. Without machine learning, a special parser generates a parse tree that represents the source logic and a fixed-size correctly encoded tree that represents the parse tree. For finetuning the logic encoder, an encoded tree generator is an artificial neural network that accepts the fixed-size encoded logic as input and responsively infers a fixed-size incorrectly encoded tree that represents the parse tree. The neural weights of the logic encoder (and optionally of the encoded tree generator) are adjusted based on backpropagation of error (i.e. loss) as a numerically measured difference between the fixed-size incorrectly encoded tree and the fixed-size correctly encoded tree.

Training syntax-aware language models with AST path prediction

Here is message compression using schema inference for condensing semantic content by removal of syntactic structure in multiple kinds of content. Topological structures of different trees are generalized to generate a merged tree. Because compression discards redundant content and often only semantic content is retained, the signal-to-noise ratio is increased, which increases accuracy of downstream semantic analytics such as machine learning. Compression based on the merged tree removes redundant information from new messages that, without obscuring semantic content, decreases the data volume for downstream analytics or archiving. This compression extracts semantic values that can be assembled into a sequence of lexical tokens that is suitable for natural language processing (NLP), and the sequence of lexical tokens does not contain tokens that represent syntax or structure. Thus, compression provides fewer tokens to be processed by a downstream language model, which is suitable for efficient processing of a live data stream.

Payload size reduction and reassembly of recursively parsable structures

Herein are machine learning techniques that adjust reconstruction loss of a reconstructive model such as an autoencoder based on importances of values of features. In an embodiment and before, during, or after training, the reconstructive model that more or less accurately reconstructs its input, a computer measures, for each distinct value of each feature, a respective importance that is not based on the reconstructive model. For example, importance may be based solely on a training corpus. For each feature during or after training, a respective original loss from the reconstructive model measures a difference between a value of the feature in an input and a reconstructed value of the feature generated by the reconstructive model. For each feature, the respective importance of the input value of the feature is applied to the respective original loss to generate a respective weighted loss. The weighted losses of the features of the input are collectively detected as anomalous or non-anomalous.

Balancing feature distributions using an importance factor

Herein is a machine learning (ML) explainability (MLX) approach in which a natural language explanation is generated based on analysis of a parse tree such as for a suspicious database query or web browser JavaScript. In an embodiment, a computer selects, based on a respective relevance score for each non-leaf node in a parse tree of a statement, a relevant subset of non-leaf nodes. The non-leaf nodes are grouped in the parse tree into groups that represent respective portions of the statement. Based on a relevant subset of the groups that contain at least one non-leaf node in the relevant subset of non-leaf nodes, a natural language explanation of why the statement is anomalous is generated.

Textual explanations for abstract syntax trees with scored nodes

Techniques are described that extend supervised machine-learning algorithms for use with semi-supervised training. Random labels are assigned to unlabeled training data, and the data is split into k partitions. During a label-training iteration, each of these k partitions is combined with the labeled training data, and the combination is used train a single instance of the machine-learning model. Each of these trained models are then used to predict labels for data points in the k−1 partitions of previously-unlabeled training data that were not used to train of the model. Thus, every data point in the previously-unlabeled training data obtains k−1 predicted labels. For each data point, these labels are aggregated to obtain a composite label prediction for the data point. After the labels are determined via one or more label-training iterations, a machine-learning model is trained on data with the resulting composite label predictions and on the labeled data set.

Generalized expectation maximization for semi-supervised learning

A method and one or more non-transitory storage media are provided to train and implement a one-hot encoder. During a training phase, computation of an encoder state is performed by executing a set of relational statements to extract unique categories in a first training data set, associate each unique category with a unique index, and generate a one-hot encoding for each unique category. The set of relational statements are executed by a query optimization engine. Execution of the set of relational statements is postponed until a result of each relational statement is needed, and the query optimization engine implements one or more optimizations when executing the set of relational statements. During an encoding phase, a set of categorical features in a second training data set are encoded based on the encoder state to form a set of encoded categorical features.

One-hot encoder using lazy evaluation of relational statements

Techniques for implementing a semi-supervised framework for purpose-oriented anomaly detection are provided. In one technique, a data item in inputted into an unsupervised anomaly detection model, which generates first output. Based on the first output, it is determined whether the data item represents an anomaly. In response to determining that the data item represents an anomaly, the data item is inputted into a supervised classification model, which generates second output that indicates whether the data item is unknown. In response to determining that the data item is unknown, a training instance is generated based on the data item. The supervised classification model is updated based on the training instance.

Semi-supervised framework for purpose-oriented anomaly detection

Herein is resource-constrained feature enrichment for analysis of parse trees such as suspicious database queries. In an embodiment, a computer receives a parse tree that contains many tree nodes. Each tree node is associated with a respective production rule that was used to generate the tree node. Extracted from the parse tree are many sequences of production rules having respective sequence lengths that satisfy a length constraint that accepts at least one fixed length that is greater than two. Each extracted sequence of production rules consists of respective production rules of a sequence of tree nodes in a respective directed tree path of the parse tree having a path length that satisfies that same length constraint. Based on the extracted sequences of production rules, a machine learning model generates an inference. In a bag of rules data structure, the extracted sequences of production rules are aggregated by distinct sequence and duplicates are counted.

Generalized production rules—N-gram feature extraction from abstract syntax trees (AST) for code vectorization

Techniques are provided for sparse ensembling of unsupervised machine learning models. In an embodiment, the proposed architecture is composed of multiple unsupervised machine learning models that each produce a score as output and a gating network that analyzes the inputs and outputs of the unsupervised machine learning models to select an optimal ensemble of unsupervised machine learning models. The gating network is trained to choose a minimal number of the multiple unsupervised machine learning models whose scores are combined to create a final score that matches or closely resembles a final score that is computed using all the scores of the multiple unsupervised machine learning models.

Sparse ensembling of unsupervised models

Unsorted sparse dictionary encodings are transformed into unsorted-dense or sorted-dense dictionary encodings. Sparse domain codes have large gaps between codes that are adjacent in order. Unlike spare codes, dense codes have smaller gaps between adjacent codes; consecutive codes are dense codes that have no gaps between adjacent codes. The techniques described herein are relational approaches that may be used to generate sparse composite codes and sorted codes.

Relational method for transforming unsorted sparse dictionary encodings into unsorted-dense or sorted-dense dictionary encodings

Techniques are described herein for using machine learning to learn vector representations of DNS requests such that the resulting embeddings represent the semantics of the DNS requests as a whole. Techniques described herein perform pre-processing of tokenized DNS request strings in which hashes, which are long and relatively random strings of characters, are detected in DNS request strings and each detected hash token is replaced with a placeholder token. A vectorizing ML model is trained using the pre-processed training dataset in which hash tokens have been replaced. Embeddings for the DNS tokens are derived from an intermediate layer of the vectorizing ML model. The encoding application creates final vector representations for each DNS request string by generating a weighted summation of the embeddings of all of the tokens in the DNS request string. Because of hash replacement, the resulting DNS request embeddings reflect semantics of the hashes as a group.

Machine learning-based DNS request string representation with hash replacement

Herein are feature extraction mechanisms that receive parsed log messages as inputs and transform them into numerical feature vectors for machine learning models (MLMs). In an embodiment, a computer extracts fields from a log message. Each field specifies a name, a text value, and a type. For each field, a field transformer for the field is dynamically selected based the field's name and/or the field's type. The field transformer converts the field's text value into a value of the field's type. A feature encoder for the value of the field's type is dynamically selected based on the field's type and/or a range of the field's values that occur in a training corpus of an MLM. From the feature encoder, an encoding of the value of the field's typed is stored into a feature vector. Based on the MLM and the feature vector, the log message is detected as anomalous.

Multi-stage feature extraction for effective ML-based anomaly detection on structured log data

Embodiments use a hierarchy of machine learning models to predict datacenter behavior at multiple hardware levels of a datacenter without accessing operating system generated hardware utilization information. The accuracy of higher-level models in the hierarchy of models is increased by including, as input to the higher-level models, hardware utilization predictions from lower-level models. The hierarchy of models includes: server utilization models and workload/OS prediction models that produce predictions at a server device-level of a datacenter; and also top-of-rack switch models and backbone switch models that produce predictions at higher levels of the datacenter. These models receive, as input, hardware utilization information from non-OS sources. Based on datacenter-level network utilization predictions from the hierarchy of models, the datacenter automatically configures its hardware to avoid any predicted over-utilization of hardware in the datacenter. Also, the predictions from the hierarchy of models can be used to detect anomalies of datacenter hardware behavior.

Datacenter level utilization prediction without operating system involvement

Herein are machine learning (ML) feature processing and analytic techniques to detect anomalies in parse trees of logic statements, database queries, logic scripts, compilation units of general-purpose programing language, extensible markup language (XML), JavaScript object notation (JSON), and document object models (DOM). In an embodiment, a computer identifies an operational trace that contains multiple parse trees. Values of explicit features are generated from a single respective parse tree of the multiple parse trees of the operational trace. Values of implicit features are generated from more than one respective parse tree of the multiple parse trees of the operational trace. The explicit and implicit features are stored into a same feature vector. With the feature vector as input, an ML model detects whether or not the operational trace is anomalous, based on the explicit features of each parse tree of the operational trace and the implicit features of multiple parse trees of the operational trace.

Extraction from trees at scale

Techniques are described herein for predicting disk drive failure using a machine learning model. The framework involves receiving disk drive sensor attributes as training data, preprocessing the training data to select a set of enhanced feature sequences, and using the enhanced feature sequences to train a machine learning model to predict disk drive failures from disk drive sensor monitoring data. Prior to the training phase, the RNN LSTM model is tuned using a set of predefined hyper-parameters. The preprocessing, which is performed during the training and evaluation phase as well as later during the prediction phase, involves using predefined values for a set of parameters to generate the set of enhanced sequences from raw sensor reading. The enhanced feature sequences are generated to maintain a desired healthy/failed disk ratio, and only use samples leading up to a last-valid-time sample in order to honor a pre-specified heads-up-period alert requirement.

Growing community of inventors

Niederweningen, Switzerland

Felix Schmidt