Intuit, Inc.

Paypal, Inc.

Itay Margolin

Yair Horesh

Tomer Handelman

Aleksandr Kim

Liran Dreval

Ofri Raviv

Shlomit Plavner

Aviv Ben Arie

Guy Shtar

Hadas Baumer

Ido Farhi

Lior Tabori

Alon Dourban

Eilon Sheetrit

Meghan Mergui

Shafik Bisharat

Yagil Ovadia

Matan Marudi

Alexsandr Kim

Shon Mendelson

Andrei Roskach

Yoni Rabin

Roy Lothan

A method including extracting a number of page features from a web page. The number of page features represent an executable logic of the web page. The method also includes embedding, by a page feature embedding model, the number of page features to generate a page vector data structure. The method also includes comparing, by a comparison model, the page vector data structure and a number of script vector data structures to identify a selected script. Each of the number of script vector data structures is generated by a script feature embedding model processing computer executable program code of a corresponding script for performing a computer function on a web page. The method also includes presenting the selected script.

Model ensemble for matching nearest script to a new script page

A method including receiving a command to perform a one-to-many matching task between first and second datasets. The first and second datasets are vectorized into first and second embedded datasets. A first self-attention model is executed on the first embedded dataset to generate a first attention dataset in which each value of a first number of first features of the first dataset is weighted based on each other value of the first number of first features. A second self-attention model is executed on the second embedded dataset to generate a second attention dataset in which each value of a second number of second features of the second dataset is weighted based on each other value of the second number of second features. The first and second attention datasets are combined into a relationship matrix expressing relationships between the first and second features. The method also includes returning the relationship matrix.

System and method for one-to-many matching between disparate datasets

Systems and methods are disclosed for training a text classification model based on long text and known semantics as training data. With a text classification model limited as to the amount of text that may be input at one time, long text that is greater than the limit may be segmented into smaller segments that are less than the limit (such as into sentences). Each segment of the long text is compared with sample segments with known impact of specific semantics to associate the long text segments with the specific semantics. To compare the long text segments with sample segments, an embedding model generates an embedding from each of the segments so that the embeddings may be compared. With the long text segments associated with specific semantics, the long text segments and the associated semantics are used as training data to train a text classification model.

Training of a text classification model from long text and based on text semantics

Certain aspects of the disclosure provide a method for detecting data collection errors by processing error data with a plurality of regression models to generate a plurality of predicted error rates over a plurality of time intervals. The method includes determining an error mode by applying a set of policy rules optimized for determining the error mode to the plurality of predicted error rates.

Real-time remote system shutdown prediction

Techniques are disclosed relating to training a machine learning model to understand one or more rules without explicitly executing the rule. In some embodiments, a computer system generates synthetic samples for a trained machine learning model usable to make a classification decision, where the synthetic samples are generated from a rule and a set of existing samples. In some embodiments, the set of existing samples are selected based on exceeding a confidence threshold for the classification decision. In some embodiments, the computer system retrains the trained machine learning model using the synthetic samples.

Training data augmentation for machine learning

Aspects of the present disclosure relate to detecting spam messages. Embodiments include creating condensed vector representations of messages; calculating similarity scores for each message relative to other messages using the vector representations; associating messages of the plurality of messages with a grouping based on the calculated similarity score for the messages within the grouping exceeding a threshold; determining that a grouping of messages are spam messages by comparing an amount of messages of the grouping of messages sent within a first time window to an amount of messages of the grouping of messages sent within a second time window, wherein the second time window comprises a time period preceding the first time window; providing the identified spam message to a machine learning model; and training the machine learning model by iteratively adjusting parameters of the model based on tracking the identified spam message through multiple layers of the model.

Spam prevention using vector-based text similarity comparisons across multiple time windows

A Large Language Model (LLM) for classifying documents by identifying indicators within the documents. A smart caching mechanism stores document classifications and associated indicators output from the LLM. The database contains document details, classifications, and associated indicators. A classification module classifies a new document by analyzing it for indicators, checking the cache for a match, and querying the database for the indicators if no match is found. The module applies a majority vote based on the classifications associated with the indicators.

Methods and systems for implementing large language models and smart caching with zero shot

Methods and systems are presented for imputing missing data items within a first dataset based on data associated with a second dataset that is the nearest neighbor of the first dataset. A first mapping model is configured to map data subsets corresponding to a first data source to first positions in a multi-dimensional space. A second mapping model is configured to map data subsets corresponding to a second data source to second positions in the multi-dimensional space. The first and second mapping models are trained together to reduce a distance between positions mapped by the first and second mapping models based on corresponding data subsets that belong to the same entity. A nearest neighbor dataset to the first dataset is identified based on the first and second mapping models. Data associated with the nearest neighbor dataset is used to impute the missing data items of the first dataset.

Deep mapping for imputing nulls

A first schema accessed is associated with a plurality of entities that are participants of an electronic system. Each entity has a corresponding attribute. Each of a first subset of the entities has a respective attribute value below a threshold. Each of a second subset of the entities has a respective attribute value above the threshold. According to the first schema, it is determined that data and/or transactions associated with the first and second subsets of the entities are processed using a first and a second model, respectively. A second schema is generated by softening the predefined threshold such that according to the second schema, data and/or transactions associated with the first subset of the entities and data and/or transactions associated with the second subset of the entities are each processed using both the first model and the second model. The second schema is implemented in the electronic system.

Hardline threshold softening

The present disclosure provides techniques for data retrieval using machine learning. One example method includes receiving a plurality of training episodes associated with different environments, wherein each training episode of the plurality of training episodes includes a sequence of states, computing, based on the plurality of training episodes, total counts of a plurality of values in the states, initializing, for each state of the sequence of states in each training episode of the plurality of training episodes, a reward based on the total counts of the plurality of values, and training a reinforcement learning agent using the rewards.

Data retrieval using machine learning

Systems and methods for training machine learning models are disclosed. An example method includes receiving a semi-labeled set of training samples including a first set of training samples, where each training sample in the first set is assigned a known label, and a second set of training samples, where each training sample in the second set has an unknown label, determining a first loss component, the first loss component providing a loss associated with the first set, determining a second loss component, the second loss component having a value which increases based on a difference between a distribution of individually predicted values of at least the second set and an expected overall distribution of at least the second set, and training the machine learning model, based on the first loss component and the second loss component, to predict labels for unlabeled input data.

Prior injections for semi-labeled samples

A processor may identify a plurality of data sets subject to upcoming update processing in a next update cycle. For each of the plurality of data sets, the processor may determine a probability that data included in the data set has changed since a most recent update processing. The processor may exclude a first subset of the plurality of data sets having respective probabilities below a threshold value from the upcoming update processing until the respective probabilities are determined again in a subsequent update cycle. The processor may perform the upcoming update processing on the plurality of the data sets not included in the first subset, where the upcoming update processing may include obtaining updated data from at least one external data source.

Efficient automatic web scraping systems and methods

A method and system that proactively generate alerts for updating a scraping script to avoid scraping script errors. A predetermined number of webpages targeted by the scraping script are randomly sampled. The scraping script is appended to each webpage in the sample. A structured list of text fragments across the webpages with the appended script is generated. At predetermined time intervals, a fresh set of webpages is sampled, the scraping script is appended to the webpages, and a new structured list is generated. If the new structured list and the previous structured list do not match, the webpages may have been changed and the scraping script may have to be updated. An alert is generated indicating that such update is required and may include a location of the mismatch. Therefore, scraping script errors are proactively detected and can be rectified before an actual error occurs and propagates.

Generating a proactive alert for outdated scraping script

Certain aspects of the present disclosure provide techniques for medoid-based data compression. One example method generally includes receiving item data indicative of one or more items, determining one or more medoids based on the item data, determining, for each item of the one or more items, a corresponding medoid based on the one or more medoids, identifying, for each item of the one or more items, a difference between the item and the corresponding medoid for the item, storing the one or more medoids, and storing, for each item of the one or more items, the identified difference between the item and the corresponding medoid.

Medoid-based data compression

Computer system security and efficiency of processing operations may be improved using techniques that are described relating to analyzing user actions based on time of day of occurrence, and using time of a day as a factor in determining whether a particular action should be allowed or disallowed. Past action times can be transformed to a two-dimensional representation using a radial time schema that avoids discontinuity. A probability distribution can indicate a likelihood of whether a new action fits a previous pattern. If a new user action is relatively unlikely due to time of day, the new user action might be denied/prevented from completing, thus enhancing computer system security and avoiding unnecessary computational processing costs.

Growing community of inventors

Tel Aviv, Israel

Itay Margolin