Все выпуски

Предлагается обобщенное семейство вероятностных тематических моделей коллекций текстовых документов, в котором эвристики регуляризации, сэмплирования, частого обновления параметров, робастности относительно шума и фона могут включаться независимо друг от друга в любых сочетаниях, порождая как известные модели PLSA, LDA, CVB0, SWB, так и новые. Показано, что робастная тематическая модель на основе PLSA, разделяющая термины на тематические, шумовые и фоновые, не нуждается в регуляризации и обеспечивает разреженность искомых дискретных распределений тем в документах и терминов в темах.

We propose a generalized probabilistic topic model of text corpora which can incorporate heuristics of Bayesian regularization, sampling, frequent parameters update, and robustness in any combinations. Wellknown models PLSA, LDA, CVB0, SWB, and many others can be considered as special cases of the proposed broad family of models. We propose the robust PLSA model and show that it is more sparse and performs better that regularized models like LDA.

Deep learning’s power stems from complex architectures; however, these can lead to overfitting, where models memorize training data and fail to generalize to unseen examples. This paper proposes a novel probabilistic approach to mitigate this issue. We introduce two key elements: Truncated Log-Uniform Prior and Truncated Log-Normal Variational Approximation, and Automatic Relevance Determination (ARD) with Bayesian Deep Neural Networks (BDNNs). Within the probabilistic framework, we employ a specially designed truncated log-uniform prior for noise. This prior acts as a regularizer, guiding the learning process towards simpler solutions and reducing overfitting. Additionally, a truncated log-normal variational approximation is used for efficient handling of the complex probability distributions inherent in deep learning models. ARD automatically identifies and removes irrelevant features or weights within a model. By integrating ARD with BDNNs, where weights have a probability distribution, we achieve a variational bound similar to the popular variational dropout technique. Dropout randomly drops neurons during training, encouraging the model not to rely heavily on any single feature. Our approach with ARD achieves similar benefits without the randomness of dropout, potentially leading to more stable training.

To evaluate our approach, we have tested the model on two datasets: the Canadian Institute For Advanced Research (CIFAR-10) for image classification and a dataset of Macroscopic Images of Wood, which is compiled from multiple macroscopic images of wood datasets. Our method is applied to established architectures like Visual Geometry Group (VGG) and Residual Network (ResNet). The results demonstrate significant improvements. The model reduced overfitting while maintaining, or even improving, the accuracy of the network’s predictions on classification tasks. This validates the effectiveness of our approach in enhancing the performance and generalization capabilities of deep learning models.

Deep learning’s power stems from complex architectures; however, these can lead to overfitting, where models memorize training data and fail to generalize to unseen examples. This paper proposes a novel probabilistic approach to mitigate this issue. We introduce two key elements: Truncated Log-Uniform Prior and Truncated Log-Normal Variational Approximation, and Automatic Relevance Determination (ARD) with Bayesian Deep Neural Networks (BDNNs). Within the probabilistic framework, we employ a specially designed truncated log-uniform prior for noise. This prior acts as a regularizer, guiding the learning process towards simpler solutions and reducing overfitting. Additionally, a truncated log-normal variational approximation is used for efficient handling of the complex probability distributions inherent in deep learning models. ARD automatically identifies and removes irrelevant features or weights within a model. By integrating ARD with BDNNs, where weights have a probability distribution, we achieve a variational bound similar to the popular variational dropout technique. Dropout randomly drops neurons during training, encouraging the model not to rely heavily on any single feature. Our approach with ARD achieves similar benefits without the randomness of dropout, potentially leading to more stable training.

To evaluate our approach, we have tested the model on two datasets: the Canadian Institute For Advanced Research (CIFAR-10) for image classification and a dataset of Macroscopic Images of Wood, which is compiled from multiple macroscopic images of wood datasets. Our method is applied to established architectures like Visual Geometry Group (VGG) and Residual Network (ResNet). The results demonstrate significant improvements. The model reduced overfitting while maintaining, or even improving, the accuracy of the network’s predictions on classification tasks. This validates the effectiveness of our approach in enhancing the performance and generalization capabilities of deep learning models.

Посты в социальных сетях способны как предсказывать движение финансового рынка, так и в некоторых случаях даже определять его направление. Анализ постов в Twitter способствует прогнозированию цен на криптовалюту. Специфика рассматриваемого сообщества заключается в особенной лексике. Так, в постах используются сленговые выражения, аббревиатуры и сокращения, наличие которых затрудняет векторизацию текстовых данных, в следствие чего рассматриваются методы предобработки такие, как лемматизация Stanza и применение регулярных выражений. В этой статье описываются простейшие модели машинного обучения, которые могут работать, несмотря на такие проблемы, как нехватка данных и короткие сроки прогнозирования. Решается задача бинарной текстовой классификации, в условиях которой слово рассматривается как элемент бинарного вектора единицы данных. Базисные слова определяются на основе частотного анализа упоминаний того или иного слова. Разметка составляется на основе свечей Binance с варьируемыми параметрами для более точного описания тренда изменения цены. В работе вводятся метрики, отражающие распределение слов в зависимости от их принадлежности к положительному или отрицательному классам. Для решения задачи классификации использовались dense-модель с подобранными при помощи Keras Tuner параметрами, логистическая регрессия, классификатор случайного леса, наивный байесовский классификатор, способный работать с малочисленной выборкой, что весьма актуально для нашей задачи, и метод k-ближайших соседей. Было проведено сравнение построенных моделей на основе метрики точности предсказанных меток. В ходе исследования было выяснено, что наилучшим подходом является использование моделей, которые предсказывают ценовые движения одной монеты. Наши модели имеют дело с постами, содержащими упоминания проекта LUNA, которого на данный момент уже не существует. Данный подход к решению бинарной классификации текстовых данных широко применяется для предсказания цены актива, тренда ее движения, что часто используется в автоматизированной торговле.

Posts on social networks can both predict the movement of the financial market, and in some cases even determine its direction. The analysis of posts on Twitter contributes to the prediction of cryptocurrency prices. The specificity of the community is represented in a special vocabulary. Thus, slang expressions and abbreviations are used in posts, the presence of which makes it difficult to vectorize text data, as a result of which preprocessing methods such as Stanza lemmatization and the use of regular expressions are considered. This paper describes created simplest machine learning models, which may work despite such problems as lack of data and short prediction timeframe. A word is considered as an element of a binary vector of a data unit in the course of the problem of binary classification solving. Basic words are determined according to the frequency analysis of mentions of a word. The markup is based on Binance candlesticks with variable parameters for a more accurate description of the trend of price changes. The paper introduces metrics that reflect the distribution of words depending on their belonging to a positive or negative classes. To solve the classification problem, we used a dense model with parameters selected by Keras Tuner, logistic regression, a random forest classifier, a naive Bayesian classifier capable of working with a small sample, which is very important for our task, and the k-nearest neighbors method. The constructed models were compared based on the accuracy metric of the predicted labels. During the investigation we recognized that the best approach is to use models which predict price movements of a single coin. Our model deals with posts that mention LUNA project, which no longer exist. This approach to solving binary classification of text data is widely used to predict the price of an asset, the trend of its movement, which is often used in automated trading.