👍🏻4 utilities for working with JSON files
Hadoop and Spark are the most popular big data frameworks for working with big data - large files. But often you need to process many small files, for example, in JSON format, which in Hadoop HDFS are distributed over many data blocks and partitions. The number of partitions determines the number of tasks since 1 task can deal with only 1 partition at a time. This will be a high load for the Application Master and reduction of productivity for the entire cluster. In addition, most of the time is spent only on opening and closing files, and not on reading data.
Therefore, it is worth to combine many small files in large one, which Hadoop and Spark can process very quickly. In the case of JSON files, such a union into an array of records can be done using the tools:
• jq – is used to filter and process incoming JSON data, great for parsing and processing data streams https://stedolan.github.io/jq/
• jo - creates JSON data structures https://github.com/jpmens/jo
• json_pp - displays JSON objects in a more convenient format, as well as convert them between textures https://github.com/deftek/json_pp
• jshon - JSON parser with fast evaluation of large amounts of data http://kmkeen.com/jshon/
https://sidk17.medium.com/boss-we-have-a-large-number-of-small-files-now-how-to-process-these-files-ee27f67dc461

👀Looking for data to train ML-models? Generate it Yourself: 3 Python Packages for Generating Synthetic Data
Synthetic data is an artificially generated, not collected, topic learning dataset for training ML models or practicing analysis techniques. You can create them yourself using possible Python packages:
• Faker is a very simple and efficient Python package for creating their data. It's great when you need to load data into a database, create a use of XML documents, prepare for load testing, or anonymize data retrieved from involved services. https://github.com/joke2k/faker
• SDV (Synthetic Data Vault) is a synthetic data storage for creating synthetic data based on a given dataset. The generated data can be a single summary, pivot table, or time series, and have the same properties and statistics as the original dataset. SDV uses synthetic data with DL models. Even if the original dataset contains multiple data types and gaps, SDV handles them. https://sdv.dev/SDV/
• Gretel Synthetics - a source code package based on a recurrent neural network for generating structured and unstructured data. The batch approach treats a data set as text data and trains a model based on it. The model will then create synthetic data with text data. Gretel is based on RNN networks, it requires more computing power, so when working with it, it is better to use Google Colab, rather than load a personal computer. https://synthetics.docs.gretel.ai/en/stable/

11 августа состоится Alfa Data Science MeetUp#2 📟

Участие бесплатное, необходимо зарегистрироваться на сайте, чтобы получить ссылку на онлайн-трансляцию.

Темы и спикеры:
🖲 Развитие клиентской базы: моделирование LTV и прогноз будущих доходов
- Сергей Королёв, Middle Data Scientist Альфа-Банк
🖲 Uplift-моделирование в ценообразовании кредитных продуктов
- Максим Коматовский, Junior Data Scientist Альфа-Банк
🖲 Совершенный код расчёт
- Максим Cтаценко, Team Lead/Senior DWH Developer в Яндекс
🖲 Побеждаем смещение распределения в задаче нейросетевого кредитного скоринга
- Алексей Фирстов, Senior Data Scientist Альфа-Банк

Митап пройдет в интерактивном формате, вопросы спикерам приветствуются, авторы лучших вопросов получат призы от Alfa Digital.

😱3 types of data anomalies
Data analysts and machine learning professionals often detect anomalies in data – detection that is not detected to pattern detection and detection. There are 3 types of anomalies:
• point anomaly, when one data point (observation) in the data set is far from the sources of the others and represents an extreme, unevenness or deviation that occurs randomly and is not related to the overall load in the data. The point anomaly also eliminates the global outlier because it is significantly different from the rest of the dataset.
• context anomaly, when a particular instance is revealed to be observable from the context. For example, in the case of time series data, such as recording a certain amount over time, the context is temporal. Data points that are very different from other data in the same sense are due to contextual outliers. For example, when the number of cars passing through the checkpoint on the border of the region in March, on average, is 1 thousand over the past 20 years. And in June, when the vacation period begins, this number reaches 8 thousand. If at the beginning of March it is 9 thousand, it will be considered an anomaly, and in the summer it will not be an anomaly. It's common for retail to see shoppers pop up during the holiday season. But a sharp increase in sales outside of holidays or sales can be called a contextual outlier.
• Collective anomaly, where a group of correlated, interrelated, or sequential instances is significantly different from the rest of the data, ie these data points are judged to be anomalous. For time series data, this may look like typical peaks and troughs occurring outside of the time period when the seasonal sequence is normal, or as a set of time series that are in outlier conditions. For example, at the same time, along with a large number of companies, there is a drop in sales, although before that there was an upward trend.
https://medium.com/datadailyread/types-of-data-anomalies-2f6fb1747eb1

3 types of data anomalies

👍🏻10 Best Practices for Naming Tables and Fields in a Database
If every developer and analyst followed these simple rules, reverse engineering would become a hobby, not a laborious job. To make it easier for you and your colleagues to work with the database, try these simple rules:
1. Separate words with underscores if the name attribute or database table consists of 2 or more words. This is a more stylish camelCase case, improved readability and conceptual platform dependency. For example, word_count.
2. Write full and semantically based names for tables and columns without reference to data types. Saving a couple of characters will do nothing but confuse. It is permissible to use benefits only where this is an abbreviated name for everyone.
3. Write the attribute name with a lowercase letter to avoid confusion from upper-case SQL keywords. It will also improve your typing speed.
4. Do not use numbers in the names of tables and columns.
5. Name the tables clearly, but briefly.
6. Name tables and columns in the singular. For example, author instead of authors.
7. Name the linking tables in alphabetical order. For example author_book
8. When an index is set, add its table and column name. For example, CREATE INDEX person_ix_first_name_last_name ON person (first_name, last_name);
9. For Boolean column type add prefix name with is_ or has_ . For example is_admin or has_membership.
10. For columns of Date-Time type, add suffix _at or _time to the name. For example, order_at or order_time.
https://dev.to/mohammadfaisal/how-to-design-a-clean-database-1e83

🔥Instead of Jupyter Notebook: Benefits of Deepnote
Jupyter notebooks have been actively using data analytics for many years and contain ML. However, despite its popularity, this research tool has significant implications:
• Difficulty in code versioning. Browsing Jupyter notebooks comes in the form of large JSON files, merging two notebooks is next to impossible. As is the usual use by developers of a Git-like tool version.
• Lack of high efficiency with IDE, code highlighting and tooltips. Usually a Data Scientist is not a professional software developer, and therefore tools that regulate the quality of the code and the value of its increase are very important.
• Difficulties of development through testing. The popular test-driven development methodology (test-driven development) is almost impossible to implement in Jupyter notebooks. Therefore, they cannot be used in data pipelines.
• Non-linear workflow due to transition from one cell to another. This may seem like an irreproducible experiment. The interactive way to encode and navigate between cells is both one of Jupyter Notebook's best features and its biggest weakness.
• Jupyter is not well suited for running long running asynchronous data scoping tasks.
😡
Many of the shortcomings are addressed as an alternative to Jupyter Notebook called Deepnote. Deepnote, like Jupyter, is an interactive notebook for solving a DS problem, but it outperforms its competitor in a number of advantages 💥:
• Real-time Collaboration - Created by Google Docs, you can share links to your notebook with colleagues, giving everyone the desired level of access (view, execute, comment, edit and full control). In addition, a cell in Deepnote allows a collaborator to leave comments, eliminating the need to switch between applications for posting messages and code for providing feedback. With access to the developer's code, managers and other board members easily follow the code development progress and development life cycle.
• Easy environment management deployment - Deepnote takes the job of installing modules and setting up the environment to run Python, including versions. In addition to Python, Deepnote also supports SQL queries.
• Deepnote has the ability to embed blocks of code in blogs and other repositories by implementing the creation of a GitHub project specifically for this purpose. Deepnote Cells allows you to embed code only, embed output only, and embed both code and output.
• Data Visualisation - Jupyter notebooks almost never evaluate EDA execution without explicit coding. Deepnote provides an early development tool in the notebook itself - the early block allows you to generate information, just like building Python, but without the need to write code.
• Save time and money - once Deepnote is in charge of code management and processing, teams don't need to commit their code pipelines to tools like GitHub, BitBucket, etc. thus reducing operating costs.
Try it for free: https://deepnote.com/

💥Instead of loops: 3 Python life hacks
Developers and data scientists know that loops in Python are slow. Instead, you can use possible alternatives:
• Map - apply a function to each value of an iterable object(list, tuple, etc. ).
• Filter - to filter out values from an iterable object (list, tuple, sets, etc.). The filtering conditions are set inside a function which is passed as an argument to the filter function.
• Reduce – this function is a bit different from the map and filter functions. It is applied iteratively to all the values of the iterable object and returns only one value.
Examples: https://medium.com/codex/3-most-efficient-yet-underutilized-functions-in-python-d865ffaca0bb

🤔Python-library to calendar operations
Python includes a built-in calendar module that includes an operation related to dates and days of the week. The functions and classes use the European calendar module, where Monday is the first day of the week and Sunday is Sunday.
To use this feature, you must first import it into your code:
import calendar
You can then call a function, for example print the names of the months in a list:
month_names = list(calendar.month_name[1:])
print(month_names)
https://docs.python.org/3/library/calendar.html

🗒Need to log Python application events? There is a special module!
Python library logging (https://docs.python.org/3/library/logging.html) defines functions and classes that implement a flexible event logging system for applications and the library. The main advantage of the logging API, an extension of this standard library, is the ability to log all events. Therefore, the Python application log can display native messages inline with messages from external modules.
The module consists of the following classes:
• Registrars require an interface that uses application code.
• Handlers send log entries (created by registrars) to the appropriate destination.
• Filters require more precise definition of the log entries to display.
• Formats for determining the location of entries in the final output.
The level of the log indicates its severity, i.e. How important is a separate message. At the basic logging level, DEBUG has the lowest priority, and CRITICAL has the highest. If we define a logger of message-sensitive logs with the DEBUG level, then all of our logged messages will be logged, since DEBUG is the lowest level. You can configure checking only for events with the ERROR and CRITICAL types.
Сode example: https://medium.com/@DavidElvis/logging-for-ml-systems-1b055005c2c2

🍁TOP-15 DS-events in September 2022 all over the world:
1. Sep 7-8 • AI for Defense Summit • Washington, DC, USA https://ai.dsigroup.org/
2. Sep 7-9 • Southern Data Science Conference 2022 • Atlanta, GA, USA https://www.southerndatascience.com/
3. Sep 12-14 • TDWI Data Governance, Quality, and Compliance. • Virtual https://tdwi.org/events/seminars/september/data-governance-quality-compliance/home.aspx
4. Sep 13-14 • Chief Data & Analytics Officers, Brazil • Brazil https://cdao-brazil.coriniumintelligence.com/
5. Sep 13-14 • Edge AI Summit • Santa Clara, CA, USA https://edgeaisummit.com/events/edge-ai-summit
6. Sep 13-15 • AI Hardware Summit • Santa Clara, CA, USA https://www.aihardwaresummit.com/events/aihardwaresummit
7. Sep 14-15 • Deep Learning Summit • London, UK https://www.re-work.co/events/deep-learning-summit-london-2022
8. Sep 14-15 • AI in Retail Summit • London, UK https://www.re-work.co/events/ai-in-retail-summit-london-2022
9. Sep 14-15 • Conversational AI Summit • London, UK https://www.re-work.co/events/conversational-ai-summit-london-2022
10. Sep 15-16 • The Modern Data Stack Conference • San Francisco, CA, USA https://www.moderndatastackconference.com/
11. Sep 21-22 • Big Data LDN • London, UK https://bigdataldn.com/
12. Sep 22 • EM Biotech Connect 2022 • Boston, MA, USA https://elementalmachines.com/em-biotech-connect-2022-0
13. Sep 22 • data.world Summit • Virtual https://data.world/events/summit/
14. Sep 26-30 • SIAM Conference on Mathematics of Data Science (MDS22) • San Diego, CA, USA https://www.siam.org/conferences/cm/conference/mds22
15. Sep 29 • Data2030 Summit 2022 • Stockholm, Sweden + Virtual https://data2030summit.com

🖕🏻3 ways to use the assignment operator in Python and a couple of reasons not to
The walrus operato (:=) to assign value provides following advantages:
• Reducing the number of function calls, for example, result = [y := func(x), y**2, y**3] instead of result = [func(x), func(x)**2, func(x) **3]
• Reducing nested conditionals, such as when typing regular expressions, by removing nested conditions if
• Simplify while loops, such as when reading files line by line or when receiving data from a socket. Instead of a dummy infinite while loop with flow control delegated to the break statement, you can use an assignment statement to reassign the value of the command, and then apply it in that conditional while loop in the same sequence, which results in significantly shorter code.
Of course, there are limitations to using this operator. For example, it is not recommended to use it with, because. when working with ContextManager(), the context is bound to the returned discussion, i.e. to the result of this method. This may be a requirement when debugging.
Also, parenthesize when mastering to ensure that the result is applied when mastered, otherwise the calculation could be done in an arbitrary order.
Code examples: https://betterprogramming.pub/should-you-be-using-pythons-walrus-operator-yes-and-here-s-why-36297be16907

Без английского в IT никак! 💻

Если ты знаешь английский, то уже наполовину знаешь IT.

Канал Английский для IT поможет тебе в этом. Тысячи слов и словосочетаний на английском, которые пригодятся тебе в изучении языков программирования.

Присоединяйся, чтобы погрузиться в компьютерный мир уже с багажом знаний! 😉

🚀Need a visual dashboard quickly? Try Panel!
Every data analyst knows that the dashboard should summarize the most important indicators for decision makers. On the one hand, the dashboard is a simple HTML page for quickly viewing graphs and text. On the other hand, making it visual and not overloaded is not so easy. Moreover, not every BI specialist has design taste and skills in working with HTML components and their interaction with Javanoscript. But data specialists actively use Python.
To create the dashboard using only Python you can with Panel - the Python open source library. It allows to create custom interactive web applications and dashboards, widgets that users connect to graphs, images, tables or text. With it, you do not need to know how to create HTML components and their interaction with Javanoscript, because it is written on Python.
Panel supports almost all plot library builds and works just as well in Jupyter notebooks as it does on a standalone secure web server. The Panel uses the same code, supports both Python and adapts HTML/JavaScript, exports applications and can export applications, and can run rich interactive applications without tying domain-specific code to any possible GUI or web interface tool.
Panel makes easy the following:
• use of Python tools for data analysis and processing:
• development in IDE or laptop environment with further deployment of applications;
• Rapid prototyping of applications and dashboards, several polished templates for final deployment;
• deep interconnection, interaction switching and events on the client side in Python;
• transfer of large and small data to an external interface;
• Authentication in the application using built-in OAuth browsers.
Usage example: https://medium.com/@jairotunior/advanced-interactive-dashboards-in-python-cc2927dcde07

🤔PandaSQL: Python Combo for Data Scientist
SQL and Pandas are the most popular data analytics tools for tabular data management, processing and analysis. They can be used independently or together in the PandaSQL, the Python-package which provides SQL syntax capabilities in the Python environment. PandaSQL allows you to query date frames Pandas uses SQL syntax.
PandaSQL is a great tool for those who know SQL and are not familiar with the Python syntax that requires Pandas. For example, in Pandas, filtering a dataframe or grouping by column column when aggregating multiple columns can be explored in a confusing way, unlike SQL. However, this simple use of SQL in Pandas has been tainted with redundant runtime. For example, to calculate the number of rows, PandaSQL requires almost 100 times more execution runtime than Pandas.
Also, Python already has a lot of names that are reserved as basic words like for, while, in, if, else, elif, import, as. SQL adds more keywords: create, like, where, having.
Thus, it is worth to try PandaSQL, but this interesting tool is not suitable for an production data analytics pipelines.
Usage examples and runtime comparison: https://towardsdatascience.com/the-downsides-of-pandasql-that-no-one-talks-about-9b63c664bef4

Visual ETL with VDP
VDP (Visual Data Preparation) is an open source visual data ETL tool for optimizing the end-to-end visual data processing pipeline. It involves extracting unstructured visual data from pre-built data sources such as cloud/local storage or IoT devices, transforming it into parsable structured data using Vision AI models, and loading the processed data into repositories, applications, or other destinations.
VDP streamlines the end-to-end visual data processing pipeline by eliminating the need for developers to create their own connectors, model service platforms, and ELT automation tools. With VDP, visual data integration becomes easier and faster. The VDP is released under the Apache 2.0 license and is available for local and cloud deployment on Kubernetes. VDP is built from a data management perspective to optimize the end-to-end flow of visual data with a transformation component that can import Vision AI models from different sources in a flexible way. Building an ETL pipeline becomes like assembling from ready-made blocks, as in a children's constructor. And high performance is provided by a Go backend with Triton Inference Server with powerful NVIDIA GPU architectures supporting TensorRT, PyTorch, TensorFlow, ONNX and Python.
VDP is also in line with MLOps, allowing one-click import and deployment of ML/DL models from GitHub, Hugging Face, or cloud storage managed by version control tools such as DVC or ArtiVC. CV Task's standardized output formats simplify data warehousing, while pre-built ETL connectors provide advanced data access through integration with Airbyte.
VDP supports different usage scenarios: synchronous for real-time inference and asynchronous for on-demand workload. The scalable API-based microservice design is developer-friendly through seamless integration with the modern data stack. NoCode/Low Code interfaces lower the barrier to entry into the technology, giving the Data Scientist and analyst independence from data engineering.
https://github.com/instill-ai/vdp

🥒7 Reasons Not to Use Pickle to Save ML Models
A Data Scientist often writes code in notebooks like Jupyter Notebook, Google Colab, or specialized IDEs. To port this code to a production environment, it must be converted to a lightweight interchange format, compressed and serialized, that is independent of the development language. One of these formats is Pickle, a binary version of a Python object for serializing and deserializing its structure, converting a hierarchy of Python objects into a stream of bytes and vice versa. The Pickle format is quite popular due to its lightness. It does not require a data schema and is quite common, but it has a number of disadvantages:
• Unsafe. You can only unpack pickle files that you trust. An attacker can create malicious data that will execute arbitrary code during decompression. You can mitigate this risk by signing the data with hmac to make sure it hasn't been tampered with. The unreliability comes not from the fact that Pickles contain code, but from the fact that they create objects by calling the constructors mentioned in the file. Any callable object can be used instead of a class name to create objects. Malicious code will use other Python callables as constructors.
• Code mismatch. If the code changes between the time the ML model is packaged in the Pickle file and when it is used, the objects may not match the code. They will still have the structure created by the old code, but will try to work with the new version. For example, if an attribute was added after the Pickle was created, the objects in the Pickle file will not have that attribute. And if the new version of the code is supposed to process it, there will be problems.
• Implicit serialization. On the one hand, the Pickle format is convenient in that it serializes any structure of a Python object. But at the same time, there is no way to specify preferences for serialization of one or another type of data. Pickle serializes everything in objects, even data that doesn't need to be serialized. But there is no way to skip the serialization of this or that attribute in Pickle. If an object contains an attribute that cannot be boxed, such as an object with an open file, Pickle will not skip it, insisting on trying to box it, and then throw an exception.
• Lack of initialization. Pickle stores the entire structure of objects. When the Pickle module recreates the objects, it does not call the init method because the object has already been created, considering the initialization to have been called when the object was first created during the creation of the Pickle file. But the init method can do some important things, like opening file objects. In this case, raw objects will be in a state that is incompatible with the init method. Or initialization can register information about the object being created. Then unselected objects will not be displayed in the general log.
• Unreadable. Pickle is a stream of binary data, i.e. instructions for the abstract execution mechanism. Once a Pickle is opened as a normal file, its contents cannot be read. To find out what is in it, you will have to use the Pickle module to load. This can make debugging difficult, since it is difficult to find the desired data in binaries.
• Binding to Python. Being a Python library, Pickle is specific to this programming language. Although the format itself can be used for other programming languages, it is difficult to find packages that provide such capabilities. Also, they will be limited to cross-language common list/dict object structures. Pickle serializes objects containing callable functions and classes without problems. But the format does not store the code, but only the name of the function or class. When unpacking data, function names are used to look for existing code in the running process.
• Low speed. Finally, compared to other serialization methods, Pickle is much slower.

🌴🌳🌲Decision Trees: brief overview
In general, decision trees are constructed via an algorithmic approach that identifies ways to split a data set based on various conditions. It is one of the most widely used and practical methods of non-parametric supervised learning used for classification and regression tasks. The goal is to create a model that predicts the value of a target variable by learning simple decision rules inferred from the data features. It is one of the most powerful and popular algorithms. Decision-tree algorithm falls under the category of supervised learning algorithms. It works for both continuous as well as categorical output variables. It learns from simple decision rules using the various data features.
Entropy is the measure of uncertainty or randomness in a data set. Entropy handles how a decision tree splits the data. The information gain measures the decrease in entropy after the data set is split. The Gini Index is used to determine the correct variable for splitting nodes. It measures how often a randomly chosen variable would be incorrectly identified. The root node is always the top node of a decision tree. It represents the entire population or data sample, and it can be further divided into different sets. Decision nodes are subnodes that can be split into different subnodes; they contain at least two branches. A leaf node in a decision tree carries the final results. These nodes, which are also known as terminal nodes, cannot be split any further.
Decision Tree Applications:
• to determine whether an applicant is likely to default on a loan.
• to determine the odds of an individual developing a specific disease.
• to find customer churn rates
• to predict whether a consumer is likely to purchase a specific product.
Advantages of Using Decision Trees
• Decision trees are simple to understand, interpret, and visualize
• They can effectively handle both numerical and categorical data
• They can determine the worst, best, and expected values for several scenarios
• Decision trees require little data preparation and data normalization
• They perform well, even if the actual model violates the assumptions
Disadvantages
• Overfitting is one of the practical difficulties for decision tree models. It happens when the learning algorithm continues developing hypotheses that reduce the training set error but at the cost of increasing test set error. But this issue can be resolved by pruning and setting constraints on the model parameters.
• Decision trees cannot be used well with continuous numerical variables.
• A small change in the data tends to cause a big difference in the tree structure, which causes instability.
https://blog.devgenius.io/decision-tree-regression-in-machine-learning-3ea6c734eb51