<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Alex Ianovski on Medium]]></title>
        <description><![CDATA[Stories by Alex Ianovski on Medium]]></description>
        <link>https://medium.com/@alexandre3k?source=rss-43fe31977190------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*1qcqo9cHptieeI0wXYDibg.jpeg</url>
            <title>Stories by Alex Ianovski on Medium</title>
            <link>https://medium.com/@alexandre3k?source=rss-43fe31977190------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 10:49:52 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@alexandre3k/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Contactless Health Insurance Cards: An Approach to Safeguard Healthcare Users and Frontline Admin]]></title>
            <link>https://medium.com/analytics-vidhya/contactless-health-insurance-cards-an-approach-to-safeguard-healthcare-users-and-frontline-admin-638e484b7043?source=rss-43fe31977190------2</link>
            <guid isPermaLink="false">https://medium.com/p/638e484b7043</guid>
            <category><![CDATA[image-processing]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[health]]></category>
            <category><![CDATA[covid19]]></category>
            <category><![CDATA[engineering]]></category>
            <dc:creator><![CDATA[Alex Ianovski]]></dc:creator>
            <pubDate>Wed, 29 Apr 2020 11:36:40 GMT</pubDate>
            <atom:updated>2020-04-30T20:37:17.282Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fzvEr1WRXEXTfe9S.png" /><figcaption>Image source: <a href="https://norbram.ca/the-new-ohip-ontario-ohip">https://norbram.ca/the-new-ohip-ontario-ohip</a></figcaption></figure><p>Contactless payment, a technology that facilitates financial transactions with the use of radio-frequency identification (RFID) or near field communication (NFC), was first introduced to Android devices in 2011 and iOS devices in 2014. As customers discovered the ease and efficiency of the payment method, adoption of the technology <a href="https://www.payments.ca/sites/default/files/canadianpaymentmethodsandtrendsreport_2019.pdf">steadily grew</a>. In March 2020, contactless payments reached an all-time <a href="https://newsroom.interac.ca/contactless-payments-during-covid-19/">utilization peak</a>. Shoppers responded to the COVID-19 outbreak by purchasing in bulk, meanwhile, retailers implemented new measures, including limiting cash-accepting lanes and enabling additional contactless payment machines to <a href="https://www.telegraph.co.uk/news/2020/03/02/exclusive-dirty-banknotes-may-spreading-coronavirus-world-health/">safeguard customers and essential frontline staff</a> from the spread of infectious disease.</p><p>Throughout Canadian healthcare institutions, a lack of NFC-enabled or digital health insurance identification means that many frontline administrative staff must currently record patient information manually. This increased interaction between frontline staff and patients poses additional health risks to both parties. In this article, I describe and demonstrate a tool that can be used to enable contactless scanning of Ontario Health Insurance Plan (OHIP) cards, introducing additional administrative efficiency and safety measures. Through the use of Amazon Textract, a service which automatically detects and extracts text and data from scanned documents, the tool can decrease interaction between frontline health care adminstators and patients, thus reducing unnecessary exposure and health risks for both parties.</p><h3>Uploading and Analyzing Image</h3><p>To start, a sample OHIP card was downloaded from an Ontario government <a href="https://news.ontario.ca/mgs/en/2016/06/gender-on-health-cards-and-drivers-licences.html">website</a> and uploaded to the Amazon S3 cloud storage service.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j5Wshd5rQvvAcgkJOrV6eg.png" /><figcaption><strong>Figure 1 | </strong>Overview of an OHIP card in Amazon S3</figcaption></figure><p>Next, the image was uploaded from S3 to Amazon Textract. Textract simplifies optical character recognition (OCR) and enables developers to launch accurate data extraction applications with ease. Textract analyzed the image for text, forms, and tables, and defined each object as a <em>Block</em>. Users can familiarize themselves with the Textract service by manually uploading documents to the <a href="https://aws.amazon.com/textract/">online console</a>. As illustrated in Figure 2 below, the background of the image is greyed out and each <em>Block </em>is highlighted in the foreground.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4plZGXhLDsJSGHJKhdt2TQ.png" /><figcaption><strong>Figure 2 </strong>| Demonstration of Amazon Textract’s optical character recognition (OCR) on an OHIP card</figcaption></figure><p>Textract’s API returned more details than the demo tool shown above. Apart from identifying objects such as words and tables in an image, the API calls returned some of the following information:</p><ul><li>Type of block obtained (e.g. key value set, page, line, word, table, cell)</li><li>Confidence level (0–100)</li><li>Identified text</li><li>Table information (row index, column index, row span, column span)</li><li>Object coordinates in image</li><li>Block dimensions</li></ul><p>Once the image was retrieved from the S3 bucket, it was converted to binary and passed to the Textract client. Identified words with a confidence level greater than 70% were then appended to an array to form a string of text.</p><pre>def get_text_analysis(self,bucket,document):<br>  # Get the document from S3<br>  print(&quot;Importing Image..........&quot;)<br>  s3_object = self.s3_resource.Object(bucket,document)<br>  s3_response = s3_object.get()<br>  stream = io.BytesIO(s3_response[&#39;Body&#39;].read())<br>  image=Image.open(stream)<br>  image_binary = stream.getvalue()<br>  print(&#39;Extracting words..........&#39;)<br>  response = self.textract_client.analyze_document(Document=   <br>   {&#39;Bytes&#39;: image_binary},FeatureTypes=[&quot;TABLES&quot;, &quot;FORMS&quot;])<br>  <br>  # Get the text blocks<br>  blocks=response[&#39;Blocks&#39;]<br>  print (&#39;Detected Document Text&#39;)<br>  <br>  # Filter words with confidence level &gt; 70%<br>  top_words = self.filter_top_words(blocks)<br>  return top_words</pre><pre># Filter word blocks with confidence level &gt; 70%<br>def filter_top_words(self,blocks):<br>   print(&quot;Filtering top words...........&quot;)<br>   top_words = []<br>   for block in blocks:<br>     if &#39;Confidence&#39; in block:<br>        print(&quot;[DEBUG] confidence = {} \ntext = {}&quot;.format(block[&#39;Confidence&#39;],block[&#39;Text&#39;]))<br>        if(block[&#39;Confidence&#39;]&gt; 70 and block[&#39;Text&#39;] not in top_words): # Remove repeated words<br>             top_words.append(block[&#39;Text&#39;])</pre><pre>    return top_words</pre><p>After filtering out low-confidence objects, such as the <em>signature</em> field and very small text, the combined output text read:</p><pre>&quot;ntarto Healt Sante ANITA JEAN WALKER 5584 - 486 - 674 674-YM - YM BORN/NE(E 1981 12-15 YRIAN MOMM DAU ag wathee ISS/DEL EXP/EXP 2012-12 - 12 - 15 20170.12-.15 201.7 <br>- 12-.15 YRAN MOM DA MOM DAA Healt Sante ANITA JEAN WALKER 5584 - 486 674 674-YM YM 1981 12-15 MOMM DAU ag wathee 2012-12 12 15 20170.12-.15 201.7 12-.15 MOM DA&quot;</pre><p>The next step required parsing the text to extract the patient’s information. Regular expressions based on expected fields such as <em>“BORN”</em>, referring to the birth date, and <em>“EXP”</em>, referring to the expiry date, were used:</p><pre>def get_ohip_dict(self,text):<br>   name = re.compile(&#39;(?&lt;=[Health] Sante )(.*?)(?= \d)&#39;).findall(text)[0]<br>   ohip_num = re.compile(&quot;(\d{4}(( |-){1,3}\d{3})( |-){1,3}\d{3})&quot;).findall(text)[0][0]<br>   ohip_chars = re.compile(&quot;[\D]{2}(?= BORN)&quot;).findall(text)[0]<br>   ohip = self.ohip_builder(ohip_num,ohip_chars)<br>   birth_year = int(re.compile(&quot;(?&lt;=BORN\/NE\(E )\d{4}&quot;).findall(text)[0])<br>   birth_month = int(re.compile(&quot;(?&lt;=BORN\/NE\(E \d{4}[ -])\d{2}&quot;).findall(text)[0])<br>   birth_day = int(re.compile(&quot;(?&lt;=BORN\/NE\(E \d{4}[ -]\d{2}[ -])\d{2}&quot;).findall(text)[0])<br>   birthdate = datetime(birth_year,birth_month,birth_day)<br>   issue_year = int(re.compile(&quot;(?&lt;=EXP\/EXP )\d{4}&quot;).findall(text)[0])<br>   issue_month = int(re.compile(&quot;(?&lt;=EXP\/EXP \d{4}-)\d{2}&quot;).findall(text)[0])<br>   issue_day = int(re.compile(&quot;(?&lt;=EXP\/EXP \d{4}-\d{2}[ -]{3})\d{2}&quot;).findall(text)[0])<br>   issuedate = datetime(issue_year,issue_month,issue_day)<br>   expdate = datetime(issue_year+5,birth_month,birth_day)<br>   ohip_dict={<br>      &quot;name&quot;: name,<br>      &quot;ohip&quot;: ohip,<br>      &quot;birthdate&quot;: birthdate,<br>      &quot;issuedate&quot;: issuedate,<br>      &quot;expdate&quot;: expdate<br>   }<br>   return(ohip_dict)</pre><p>Finally, the fields of interest were stored in a dictionary object. The object could to be later retrieved or uploaded to a database.</p><pre><strong>Table 1</strong> | Resulting regions of interest captured from Health Insurance ID<br>              <br>              || <strong>Property       </strong>| <strong>Value               </strong>||<br>              || name           | &#39;ANITA JEAN WALKER&#39; ||<br>              || ohip           | &#39;5584486674YM&#39;      ||<br>              || birthdate      | 15/12/1981          ||<br>              || issuedate      | 12/12/2012          ||<br>              || expdate        | 15/12/2017          ||</pre><p>As shown in Table 1 above, 98% of the characters of interest were extracted accurately. Accuracy of results could be further improved by measures such as increasing the sample size and capturing multiple images for each ID card, increasing the image resolution, and utilizing the Block coordinates to search for expected fields located in spatial regions on the identification card.</p><p>Mobile digital wallet applications could also be utilized to store health insurance identification cards and enable contactless identification. The province of Ontario has worked with <a href="https://bluink.ca/eid-me">Bluink</a> to digitize and securely store Ontario driver’s licenses. Expanding this infrastructure to digitize OHIP could reduce the interaction between administrative staff and patients, reducing the risk of spreading infectious disease.</p><h3>Conclusion</h3><p>Limiting physical payment transactions between customers and essential frontline staff has been reported to reduce the risk of spreading infectious disease. Currently, health ministries do not deploy contactless health insurance cards, and patient information is recorded manually by frontline workers. This article demonstrates a tool for enabling contactless ID without modifying existing identification cards or implementing additional infrastructure such as chip readers. This solution can assist healthcare users and frontline administrative staff by reducing interaction time through automated digitization of health insurance information, thus mitigating risk of unnecessary exposure and contact.</p><p><em>As always, thank you for taking the time to read this material. Please share any comments and thoughts below; any and all feedback is appreciated. In addition, please feel free to reach out via LinkedIn.</em></p><p><a href="https://www.linkedin.com/in/ianovski/">Alex Ianovski - Sr. Release Engineer - Rogers Communications | LinkedIn</a></p><h3>Resources</h3><p><a href="https://github.com/ianovski/healthcard-scanner">ianovski/healthcard-scanner</a></p><h3>References</h3><ul><li><a href="https://www.telegraph.co.uk/news/2020/03/02/exclusive-dirty-banknotes-may-spreading-coronavirus-world-health/">Dirty banknotes may be spreading the coronavirus, WHO suggests</a></li><li><a href="https://www.payments.ca/industry-info/industry-data">Industry Data</a></li><li><a href="https://newsroom.interac.ca/contactless-payments-during-covid-19/">Contactless payments during COVID-19 - Interac Reports</a></li><li><a href="https://aws.amazon.com/textract/">OCR Software, Data Extraction Tool - Amazon Textract - AWS</a></li><li><a href="https://news.ontario.ca/mgs/en/2016/06/gender-on-health-cards-and-drivers-licences.html">Gender on Health Cards and Driver&#39;s Licences</a></li><li><a href="https://norbram.ca/the-new-ohip-ontario-ohip">OHIP + - Here and Gone</a></li><li><a href="https://bluink.ca/eid-me">eID-Me: Remote ID Verification and Mobile Digital ID</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=638e484b7043" width="1" height="1" alt=""><hr><p><a href="https://medium.com/analytics-vidhya/contactless-health-insurance-cards-an-approach-to-safeguard-healthcare-users-and-frontline-admin-638e484b7043">Contactless Health Insurance Cards: An Approach to Safeguard Healthcare Users and Frontline Admin</a> was originally published in <a href="https://medium.com/analytics-vidhya">Analytics Vidhya</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Analyzing wine bottle labels with AWS Textract]]></title>
            <link>https://medium.com/@alexandre3k/analyzing-wine-bottle-labels-with-aws-textract-d255a1c9d572?source=rss-43fe31977190------2</link>
            <guid isPermaLink="false">https://medium.com/p/d255a1c9d572</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[wine]]></category>
            <category><![CDATA[image-processing]]></category>
            <category><![CDATA[recommendations]]></category>
            <category><![CDATA[aws]]></category>
            <dc:creator><![CDATA[Alex Ianovski]]></dc:creator>
            <pubDate>Sat, 21 Mar 2020 12:57:53 GMT</pubDate>
            <atom:updated>2020-04-23T01:22:53.357Z</atom:updated>
            <content:encoded><![CDATA[<h3>OCR Powered Search: Analyzing unstructured wine bottle labels with Amazon Textract</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wjsSsBTuxW_2GJ4N" /><figcaption>Photo by <a href="https://unsplash.com/@scottiewarman?utm_source=medium&amp;utm_medium=referral">Scott Warman</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Background</h3><p>In 2018, AWS announced a number of high-level machine learning (ML) tools at its re:Invent conference. Among them was <a href="https://aws.amazon.com/textract/">Textract</a>, an image processing API made for easily extracting objects such as text and tables from virtually any document or image in JPG, PNG, or PDF format.</p><figure><img alt="Example of using Textract to automatically detect key-value pairs from document images." src="https://cdn-images-1.medium.com/max/1024/1*tLi--D0Nzm7RLHV5wS8-SQ.png" /><figcaption><strong>Figure 1 </strong>| Example of using Textract to automatically detect key-value pairs from document images. Source: <a href="https://aws.amazon.com/textract/features/">https://aws.amazon.com/textract/features/</a></figcaption></figure><p>Having implemented image processing libraries OpenCV and Tesseract in previous projects, I was familiar with the effort required to tune images for stable optical character recognition (OCR). Achieving consistent results can be a headache, especially when factors such as image quality, lighting, and image dimensions cannot be easily controlled. A major benefit of implementing AWS’s OCR tool is that it does not require any prior knowledge of ML or image processing, drastically reducing development time.</p><p>Inspired by marketplace apps such as Vivino and Untappd, I aimed at developing a proof of concept application to help shoppers find a bottle of wine they would enjoy by scanning a bottle label. Wine bottle labels can be particularly challenging for OCR, as they can vary drastically in style, font, language, and information displayed.</p><figure><img alt="Screenshots of mobile application, Vivino, which enables customers to scan wine bottle and pull crowd sourced reviews." src="https://cdn-images-1.medium.com/max/1000/0*u0csgtk4QjfsOI6P.png" /><figcaption><strong>Figure 2 | </strong>Screenshots of mobile application, Vivino, which enables customers to scan wine bottles and read crowd sourced reviews. Source: <a href="https://hauteliving.com/2015/09/4-wine-apps-that-make-you-the-sommelier/585481/">https://hauteliving.com/2015/09/4-wine-apps-that-make-you-the-sommelier/585481/</a></figcaption></figure><h3>Image Processing</h3><p>An image of a wine bottle was taken on a mobile phone and then uploaded to <a href="https://aws.amazon.com/s3/">S3</a>, AWS’s cloud storage system. Files can be uploaded either <a href="https://boto3.amazonaws.com/v1/documentation/api/latest/guide/s3-uploading-files.html">programatically</a> or manually using AWS’s online management console.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*seFXV4Kk1sxiZ-c6_Q4N3A.png" /><figcaption><strong>Figure 3| </strong>Viewing files located in an S3 bucket using AWS’s online management console.</figcaption></figure><p>AWS users can familiarize themselves with the Textract tool by manually uploading documents to the <a href="https://aws.amazon.com/textract/">online console</a>. Each text, form, or table identified in the image is defined as a <em>Block </em>and a rectangle highlighting each block is superimposed on the image.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EIc8LrD6pP8zJnPZqvXZKg.png" /><figcaption><strong>Figure 4|</strong> Amazon’s online Textract demo tool extracting text from a wine bottle label.</figcaption></figure><p>As shown in <strong>Figure 4</strong>, Textract successfully identified most words regardless of font, size, or colour, without any tuning to the image processing algorithms.</p><p>Textract’s API implementation returns more details than the online tool. Apart from identifying objects such as words and tables in an image, the API returns some of the following information:</p><ul><li>Type of block obtained (e.g. key value set, page, line, word, table, cell)</li><li>Confidence level (0–100)</li><li>Identified text</li><li>Table information (row index, column index, row span, column span)</li><li>Object coordinates in image</li><li>Block dimensions</li></ul><p>The proof of concept application was written in Python 3.8. The wine label image was downloaded from an S3 bucket, converted to binary and passed to the Textract client. Identified words with a confidence level greater than 90% were appended to an array to form a search query.</p><pre>def get_text_analysis(self,bucket,document):<br>  # Get the document from S3<br>  print(&quot;Importing Image..........&quot;)<br>  s3_object = self.s3_resource.Object(bucket,document)<br>  s3_response = s3_object.get()<br>  stream = io.BytesIO(s3_response[&#39;Body&#39;].read())<br>  image=Image.open(stream)<br>  image_binary = stream.getvalue()<br>  print(&#39;Extracting words..........&#39;)<br>  response = self.textract_client.analyze_document(Document=   <br>   {&#39;Bytes&#39;: image_binary},FeatureTypes=[&quot;TABLES&quot;, &quot;FORMS&quot;])<br>  <br>  # Get the text blocks<br>  blocks=response[&#39;Blocks&#39;]<br>  print (&#39;Detected Document Text&#39;)<br>  <br>  # Filter words with confidence level &gt; 90%<br>  top_words = self.filter_top_words(blocks)<br>  return top_words</pre><h3>Querying Product Information</h3><p>After forming a search query, the application required a database containing valuable product information such as description, rating, price, grape variety, and region. <a href="https://www.programmableweb.com/category/wine/api">Several APIs</a> exist for wine databases; however, these APIs require structured queries for each input variable. A structured search query is particularly challenging for this image processing application, since wine labels vary widely in what information is displayed and the order in which it appears. For instance, in this case the wine label identified words such as “SINGLE ESTATE”, “HOYA”, and “DECADENAS”. It would be very challenging to determine which of this strings represents the wine brand. Additionally, some letters are difficult to distinguish by the OCR (e.g. ‘O’ vs. ‘G’) and resulted in minor typos which would be challenging to correct.</p><p>To avoid such difficulties, I appended the string “wine enthusiast” to the end of the unstructured query, and wrote a bot powered by Selenium to run a Google search of the results. The bot then navigated to the first result on the Wine Enthusiast website and scraped the relevant information.</p><p>For the wine label shown in <strong>Figure 2, </strong>the web scraper identified the following information:</p><pre>Description: &quot;Earthy, leathery aromas of plum and berry are a touch rustic. This feels rubbery and raw, but fresh. Dry, earthy plum flavors are limited, while the finish is short, <br>with a tight, chalky feel. MICHAEL SCHACHNER&quot;<br>Rating: &quot;Not yet rated&quot;<br>Price: 10<br>Designation: Hoya de Cadenas Reserva<br>Variety: Tempranillo<br>Appelation: Utiel-Requena<br>Winery: Vicente Gandia</pre><p>As shown above, the scraper correctly identified and printed the relevant wine information from the wine database website.</p><pre>def get_wine_info(self, search):<br>   options = Options()<br>   options.add_argument(&quot;--headless&quot;)<br>   print(&#39;Getting Wine Info..........&#39;)<br>   driver = webdriver.Chrome(executable_path=r&quot;path&quot;,<br>    options=options)<br>   driver.get(&quot;https://www.google.com&quot;)<br>   search_bar = WebDriverWait(driver,10).until(        <br>    EC.presence_of_element_located((By.XPATH, &quot;//*[@type=&#39;text&#39;]&quot;)))</pre><pre>   search_bar.send_keys(search)<br>   WebDriverWait(driver,10).until(<br>    EC.presence_of_element_located((By.XPATH, <br>    &quot;(//*[@value=&#39;Google Search&#39;])[2]&quot;))).click()</pre><pre>   WebDriverWait(driver,10).until(<br>    EC.presence_of_element_located((By.XPATH, &quot;(//h3)[1]&quot;))).click()<br>   <br>   print(&#39;Compiling Report..........&#39;)</pre><pre>   description = WebDriverWait(driver,10).until(<br>     EC.presence_of_element_located((By.XPATH,<br>     &quot;//*[@class=&#39;description&#39;]&quot;))).text</pre><pre>   rating = WebDriverWait(driver,10).until(<br>     EC.presence_of_element_located((By.XPATH, <br>     &quot;//*[@id=&#39;user-rating&#39;]&quot;))).text</pre><pre>   price = WebDriverWait(driver,10).until(<br>     EC.presence_of_element_located((By.XPATH, <br>     &quot;(//*[@class=&#39;info medium-9 columns&#39;])[1]/span/span&quot;))).text</pre><pre>   designation = WebDriverWait(driver,10).until(<br>     EC.presence_of_element_located((By.XPATH, <br>     &quot;(//*[@class=&#39;info medium-9 columns&#39;])[2]/span/span&quot;))).text</pre><pre>   variety = WebDriverWait(driver,10).until(<br>     EC.presence_of_element_located((By.XPATH, <br>     &quot;(//*[@class=&#39;info medium-9 columns&#39;])[3]/span/a&quot;))).text</pre><pre>   appelation = WebDriverWait(driver,10).until(<br>     EC.presence_of_element_located((By.XPATH, <br>     &quot;(//*[@class=&#39;info medium-9 columns&#39;])[4]/span/a&quot;))).text</pre><pre>   winery = WebDriverWait(driver,10).until(<br>    EC.presence_of_element_located((By.XPATH, <br>    &quot;(//*[@class=&#39;info medium-9 columns&#39;])[5]/span/span/a&quot;))).text</pre><pre>   price = list(map(int, re.findall(r&#39;\d+&#39;,price)))</pre><pre>   wine_info = {&#39;description&#39;:description,&#39;rating&#39;:rating,<br>    &#39;price&#39;:str(price[0]),&#39;designation&#39;:designation,<br>    &#39;variety&#39;:variety,&#39;appelation&#39;:appelation,&#39;winery&#39;:winery}</pre><pre>  return wine_info</pre><h3>Conclusion</h3><p>This article demonstrated how a simple application could be built to scan wine bottle labels and provide relevant product information to customers. Amazon Textract enabled me to quickly develop a robust image processing application without implementing any ML or image processing techniques. I was very impressed by Textract’s ease of use and am looking forward to exploring AWS’s other ML tools.</p><p>Feel free to try out the code described in this article and get in touch on LinkedIn:</p><ul><li><a href="https://github.com/ianovski/wine-analyzer">ianovski/wine-analyzer</a></li><li><a href="https://www.linkedin.com/in/ianovski/">Alex Ianovski - Sr. Release Engineer - Rogers Communications | LinkedIn</a></li></ul><h3><strong>References</strong></h3><ul><li><a href="https://aws.amazon.com/textract/features/">Amazon Textract Features | AWS</a></li><li><a href="https://hauteliving.com/2015/09/4-wine-apps-that-make-you-the-sommelier/585481/">4 Wine Apps That Make You The Sommelier</a></li><li><a href="https://aws.amazon.com/s3/">Amazon S3 - Cloud Object Storage - AWS</a></li><li><a href="https://www.programmableweb.com/category/wine/api">Wine APIs</a></li><li><a href="https://www.wineenthusiast.com/">Wine Enthusiast - Buy Wine Storage, Accessories and Wine Gifts</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d255a1c9d572" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sentiment analysis: exploring pre-trained and domain-dependent models]]></title>
            <link>https://medium.com/analytics-vidhya/sentiment-analysis-exploring-pre-trained-and-domain-dependent-models-9604b4f48561?source=rss-43fe31977190------2</link>
            <guid isPermaLink="false">https://medium.com/p/9604b4f48561</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[sentiment-analysis]]></category>
            <category><![CDATA[word2vec]]></category>
            <category><![CDATA[transfer-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Alex Ianovski]]></dc:creator>
            <pubDate>Mon, 07 Oct 2019 03:16:28 GMT</pubDate>
            <atom:updated>2019-10-11T08:26:07.000Z</atom:updated>
            <content:encoded><![CDATA[<h3>Sentiment Analysis: Exploring Pre-trained and Domain-dependent Models</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*obXIi9pJ5uyBSaGR" /><figcaption>Photo by <a href="https://unsplash.com/@franckinjapan?utm_source=medium&amp;utm_medium=referral">Franck V.</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Organizations and researchers are implementing machine learning (ML) models to assist in identifying a myriad of insights. For many organizations, this includes monitoring key performance indicators (KPI’s), quantifying user satisfaction, and identifying security threats. In my <a href="https://medium.com/@alexandre3k/understanding-customer-needs-with-machine-learning-de19afeebc08">previous article</a>, I demonstrated how to build a model to analyze customer reviews and classify sentiment as either <em>positive, negative, </em>or<em> neutral</em>. In reality, not all research teams or organizations have the <strong>resources available to build their own models. </strong>Instead, many teams prefer to use <strong>pre-trained models </strong>when conducting ML tasks such as natural language processing (NLP).</p><p><strong>This article aims to provide organizations and researchers with a decision-making resource for conducting sentiment analysis; a tool to help determine whether to use an existing pre-trained model or develop new model.</strong> This article will begin with the concept of transfer learning; it will then evaluate six popular pre-trained sentiment analysis models and one domain-dependent model, and conclude with a cost comparison of each approach.</p><h3>Transfer Learning</h3><p>Developing a model requires a<strong> large training dataset, labour for cleaning and labeling the data, and technical resources for development</strong>. For teams seeking to deploy a solution and gain insights more quickly, a <strong><em>transfer learning </em>approach is more feasible</strong>. Instead of training models on domain-specific datasets (such as your organization’s customer reviews or survey responses), transfer learning models are trained on<strong> very large, generalized corpora containing millions of data entries</strong>(e.g. Wikipedia articles or publicly accessible movie reviews). These pre-trained solutions are easily implemented and accessible through local libraries or cloud APIs. However, <strong>users must approach pre-trained solutions skeptically, validating the models and optimizing their investment. </strong>This article provides insights by evaluating and comparing a domain dependent model with the following popular pre-trained sentiment analysis models:</p><ul><li>Google Natural Language</li><li>IBM Watson Natural Language Understanding</li><li>Microsoft Azure Text Analytics</li><li>Amazon Comprehend</li><li>TextBlob Simplified Text Processing</li><li>VADER Sentiment Analysis</li></ul><h3>Model Evaluation</h3><p>15,000 customer reviews of airline companies labelled as either <em>positive, negative, </em>or <em>neutral, </em>were acquired. A subset of 12,000 reviews was used to train the <a href="https://medium.com/@alexandre3k/understanding-customer-needs-with-machine-learning-de19afeebc08">domain-dependent model</a>, and the remaining 3,000 reviews were used for evaluating the performance of all models. The accuracy, precision, recall and F1-score were calculated for each sentiment label and model, and plotted in <strong>Figure 1</strong> and <strong>Figure 2</strong> below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5yeLGKVTcWrOSjH9m0zHPA.png" /><figcaption><strong>Figure 1 </strong>| Comparison of model accuracy for labeling sentiment of customer reviews as negative, neutral, and positive.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*snJINuUX0E5FwjtpVGB_4A.png" /><figcaption><strong>Figure 2 </strong>| Comparison of precision, recall and F1-score of models labeling sentiment of customer reviews as negative, neutral, and positive.</figcaption></figure><p>The mean of each metric was calculated and shown in <strong>Table 1</strong> below. <strong>The top performing models were IBM and AWS followed by the domain-specific model. </strong>This demonstrated that models trained on large datasets from different domains could out perform a model which was trained on domain-specific data.</p><pre><strong>Table 1 | </strong>Mean Accuracy, Precision, Recall, and F1-Score of sentiment analysis models.</pre><pre>           Model         Accuracy Precision Recall F1-Score<br>           ------------- -------- --------- ------ --------<br>           Word2Vec + RF 0.72     0.57      0.66   0.60<br>           Google        0.50     0.57      0.64   0.52<br>           IBM          <strong> </strong>0.76     0.71      <strong>0.69   0.68</strong><br>           MSFT          0.61     0.59      0.53   0.52<br>           AWS           <strong>0.77     0.73</strong>      0.66   0.66<br>           TextBlob      0.34     0.49      0.54   0.36<br>           VADER         0.53     0.59      0.54   0.53</pre><h3>Costs and Resources</h3><p>When deploying a pre-trained model, cost is determined by the number of documents analyzed by the user. As evident by the results in Table 2, costs of sentiment analysis solutions vary.</p><pre><strong>Table 2 | </strong>Cost per 100,000 Sentiment Analysis API requests.</pre><pre>                        Model      Cost($)<br>                        ---------  -------<br>                        Google     100<br>                        IBM        300<br>                        Microsoft  250<br>                        AWS        10<br>                        TextBlob   0<br>                        VADER      0</pre><p>One benefit for teams building a model trained on domain-specific data is that <strong>running the model is free, without any recurring costs.</strong> However, <strong>resources are required for training the model</strong> including labour for preparing, cleaning and labeling the data, as well as technical expertise for building, tuning and evaluating the model. Resources such as Amazon Mechanical Turk (MTurk) are an economic and efficient way to outsource the tedious and time-consuming task of labeling, training, and testing datasets. For instance, labeling the sentiment of 15,000 customer reviews costs $750, which is sufficient to achieve the results shown above. Furthermore, the performance of ML algorithms improves as the size of the training dataset increases. Therefore, an additional benefit of training a model on domain-dependent data is that it will improve as more data is ingested, and potentially out performing the pre-trained solutions.</p><h3>Conclusion</h3><p>Large digitized corpora of text and advances in ML algorithms have led to significant improvements in transfer learning for NLP. These developments enable teams to benefit from the implementation ease provided by pre-trained models, and to deploy ML solutions more quickly. The solutions are particularly useful for teams with limited sized datasets, that would otherwise be insufficient to train a robust model. As highlighted above, teams must evaluate solutions from different vendors before adopting a pre-trained model, as performance varies.</p><p>For teams with access to large training datasets, the traditional approach of training a model on domain-specific data remains a priority. For example, a simple model with only 12,000 training records outperformed four out of six pre-trained models. By utilizing resources such as MTurk, teams can expedite the tedious task of labeling datasets, and instead focus their energy on developing robust models and deploying solutions sooner. An additional benefit to developing a content-dependent model is that the dataset can be re-used for performing other NLP tasks. Organizations can reuse survey results to conduct other NLP tasks such as named entity recognition to identify customer pain points related to particular products. For this task, pre-trained models may not be optimal as they will not contain the specific product names which are relevant to the organization.</p><p>It is my hope that this article has added value in an organization or individual decision-making process related to sentiment analysis or any transfer learning approach. As always, thank you for taking the time to read the material. Please share any comments and thoughts below; any and all feedback on this article is appreciated. In addition, please feel free to reach out via LinkedIn.</p><p><a href="https://www.linkedin.com/in/ianovski/">Alex Ianovski - Sr. Release Engineer - Rogers Communications | LinkedIn</a></p><h3>Resources</h3><p><a href="https://github.com/ianovski/customer-review-sentiment">ianovski/customer-review-sentiment</a></p><h3>References</h3><ul><li><a href="https://medium.com/@alexandre3k/understanding-customer-needs-with-machine-learning-de19afeebc08">Understanding customer needs with Machine Learning</a></li><li><a href="https://towardsdatascience.com/cost-comparisons-for-cloud-sentiment-analysis-at-amazon-google-ibm-and-microsoft-bbe92cb620d7">Cost Comparisons for Cloud Sentiment Analysis at Amazon, Google, IBM and Microsoft</a></li><li><a href="https://www.mturk.com/">Amazon Mechanical Turk</a></li><li><a href="https://cloud.google.com/products/ai/">Cloud AI | Google Cloud</a></li><li><a href="https://www.ibm.com/watson/services/natural-language-understanding/">Natural Language Understanding</a></li><li><a href="https://azure.microsoft.com/en-ca/services/cognitive-services/text-analytics/">Text Analytics API | Microsoft Azure</a></li><li><a href="https://docs.aws.amazon.com/comprehend/latest/dg/API_DetectSentiment.html">DetectSentiment</a></li><li><a href="https://textblob.readthedocs.io/en/dev/">TextBlob: Simplified Text Processing - TextBlob 0.19.0 documentation</a></li><li><a href="https://github.com/cjhutto/vaderSentiment">GitHub - cjhutto/vaderSentiment: VADER Sentiment Analysis. VADER (Valence Aware Dictionary and sEntiment Reasoner) is a lexicon and rule-based sentiment analysis tool that is specifically attuned to sentiments expressed in social media, and works well on texts from other domains.</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9604b4f48561" width="1" height="1" alt=""><hr><p><a href="https://medium.com/analytics-vidhya/sentiment-analysis-exploring-pre-trained-and-domain-dependent-models-9604b4f48561">Sentiment analysis: exploring pre-trained and domain-dependent models</a> was originally published in <a href="https://medium.com/analytics-vidhya">Analytics Vidhya</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding customer needs with Machine Learning]]></title>
            <link>https://medium.com/@alexandre3k/understanding-customer-needs-with-machine-learning-de19afeebc08?source=rss-43fe31977190------2</link>
            <guid isPermaLink="false">https://medium.com/p/de19afeebc08</guid>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[customer-experience]]></category>
            <category><![CDATA[customer-service]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Alex Ianovski]]></dc:creator>
            <pubDate>Sun, 28 Jul 2019 23:57:34 GMT</pubDate>
            <atom:updated>2019-07-29T15:25:03.956Z</atom:updated>
            <content:encoded><![CDATA[<h3>Understanding customer reviews with machine learning</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KaXPzzRCA7BCM7kj" /><figcaption>Photo by <a href="https://unsplash.com/@johnschno?utm_source=medium&amp;utm_medium=referral">John Schnobrich</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>In a recent episode of one of my favourite <a href="https://www.youtube.com/watch?v=JcGY6osEAv0">podcasts</a>, the guest, Ellen Loeshelle,<strong> </strong>discussed software for <strong>analyzing customer reviews and classifying sentiment on companies’ products.</strong> Why is this important? As technologies advance and industries become increasingly competitive,<strong> focus on customer experience becomes critical to maintaining loyalty and driving long-term business strategy.</strong> Companies must quickly identify where they are succeeding with customers and where there are opportunities to improve.</p><p>Inspired by the podcast episode, I set out to build my own classification model for analyzing customer reviews and classifying them as either (i) positive, (ii) negative, or (iii) neutral. However, rather than implementing a rule-based approach to classify reviewer sentiment, as proposed by Loeshelle,<strong> I used machine learning (ML) algorithms such as word2vec and random forest classification. </strong>Overall, the model successfully identified negative reviews with 92% accuracy, and achieved an overall accuracy of 71%.</p><h3><strong>Acquiring Data</strong></h3><p><strong>Now, more than ever before, conversations between businesses and customers are open dialogues. </strong>Businesses use platforms such as Facebook, Twitter and Instagram to distribute company and product information to vast audiences. Customers, in-turn, use these platforms to react, share and voice their opinions on companies and said products.<strong> For large organizations offering a wide range of products and services, reading every review from a myriad of sources including online forums, emails, social media, and call-center transcripts is not feasible</strong>. However, using Natural Language Processing (NLP) concepts and a large corpus of customer reviews, ML algorithms can be trained to automatically classify customer feedback.</p><p>For this project, I used an open dataset containing <a href="https://www.kaggle.com/crowdflower/twitter-airline-sentiment">15,000 labeled customer reviews of airline companies</a>, collected from users on Twitter.</p><p>The sentiment of the customer reviews is shown in <strong>Figure 1 </strong>below. Apart from Virgin America, customer sentiment on the airlines is imbalanced and over-represented with negative reviews. Imbalanced datasets are problematic for training ML algorithms, as there is insufficient data to model under-represented labels (in this case, ‘neutral’ and ‘positive’ reviews). In this iteration, I do not implement any techniques to deal with class distribution, however, this would be useful for improving sensitivity of the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R8lVykYvS0bHte_K.png" /><figcaption><strong>Figure 1 | </strong>Visualization of the sentiment of six airline companies by 15,000 twitter users. Source: <a href="https://www.kaggle.com/crowdflower/twitter-airline-sentiment">https://www.kaggle.com/crowdflower/twitter-airline-sentiment</a></figcaption></figure><h3>Preprocessing</h3><p>Characters which appeared in the Tweets but would be problematic for training the model were removed. This included html tags, and non-letter characters such as emojis.</p><p>Stop-words and filler-words such as “the”, “and”, “I” etc, were also removed. These words occur frequently but do not indicate positive or negative sentiment and are not useful in training the model. Following this cleaning process, the most frequently used words in the dataset are displayed in <strong>Figure 2</strong> below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4HTK1hOzwJsKQFeILD0vlg.png" /><figcaption><strong>Figure 2 | </strong>Most frequently words used in customer reviews of airlines.</figcaption></figure><h3>From words to vectors with Word2Vec</h3><p>A major challenge of analyzing free-flowing text is the inherent lack of structure. In order to utilize ML algorithms we must first convert the unstructured text into vectors. <a href="https://code.google.com/archive/p/word2vec/">Word2vec</a> is an algorithm developed by Google which <strong>converts text from a large corpus into vectors.</strong> It first constructs a vocabulary from the training text data and then learns vector representation of words. <strong>Vectors of words with similar meaning appear close together</strong> and as a result, words and the resulting word-vector model can be used as a feature in many NLP and ML applications.</p><h3>Visualizing</h3><p>Word2vec is commonly implemented with 300 dimensional vectors. Unfortunately, those are too many dimensions for people to visualize. Fortunately, we can use the t-SNE algorithm to visualize the word-vectors in 2-D.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AGIqWekw3IInMpSvBi6j4A.png" /><figcaption><strong>Figure 3 | </strong>2D t-SNE representation of word2vec model trained using airline customer reviews.</figcaption></figure><p>Examining <strong>Figure 3</strong>, we observe that clusters containing similar words are formed. The word2vec library provides the similar_by_word() method. When provided with a word, the method returns the words with the shortest cosine distance. We can see the results for some examples:</p><p><strong>Terrible:</strong> &#39;horrible&#39;,&#39;poor&#39;,&#39;received&#39;,&#39;awful&#39;,&#39;sucks&#39;,&#39;desk&#39;,&#39;bad&#39;,&#39;cust&#39;,&#39;rude&#39;</p><p><strong>Americanair:</strong> &#39;southwestair&#39;,&#39;united&#39;,&#39;virginamerica&#39;,&#39;jetblue&#39;,&#39;amp&#39;,&#39;usairways&#39;</p><p><strong>Call:</strong> &#39;phone&#39;,&#39;someone&#39;,&#39;hold&#39;,&#39;hung&#39;,&#39;back&#39;,&#39;speak&#39;,&#39;answer’,&#39;line&#39;,&#39;online&#39;</p><h3>Averaging vectors and Random Forest Classification</h3><p>Our dataset is divided such that 80% of the data will be used for training the model and the remaining 20% will be used for testing and validation. As each word within a customer review is an independent vector, we can calculate the average word-vector of each review. Thus, we can correspond a vector, representing a customer review, to a sentiment label (‘positive’, ‘neutral’, or ‘negative’). And, we are ready to train our model!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9a0FKXDZVjqEXX7oZxQUSw.png" /><figcaption><strong>Figure 4 </strong>| A) Normalized confusion matrix for validating the random forest model. B) Comparison of actual and predicted results.</figcaption></figure><p>Using a random forest classification model with 100 trees, we achieve an accuracy of 92%, 38%, and 41% for predicting <em>negative</em>, <em>neutral</em>, and <em>positive </em>reviews. The recall for the above classes was 92%, 38%, and 43%, respectively. The precision was 77%, 56%, and 70%, respectively. The higher accuracy of classifying negative reviews was likely caused by the imbalance of sentiment among the reviews. Furthermore, word2vec performs best when trained on datasets containing billions of words, and thus better results could be expected using a larger dataset of labelled customer reviews.</p><h3>Future Steps</h3><p>I look forward to developing this project further and documenting my progress.</p><p>Next, I will evaluate the techniques presented here against pre-trained sentiment analysis models. Although pre-trained models are not trained on domain-specific data, the transfer-learning approach benefits from very large corpora and combinations of complex models and techniques. Pre-trained models are also readily available and simple to implement, making them attractive to corporations that do not have the resources to develop in-house solutions.</p><p>Additionally, I plan to implement <em>named entity recognition</em> to identify the subject of each review. This could provide insight into sentiment around specific products and services and generate suggestions to businesses on areas for improvement. I have not reviewed the literature on this topic yet and welcome any suggestions on how to approach this problem!</p><p>Thank you for reading. Please share any comments and thoughts below, or get in touch via LinkedIn.</p><p><a href="https://www.linkedin.com/in/ianovski/">Alex Ianovski - Sr. Release Engineer - Rogers Communications | LinkedIn</a></p><h3>Resources</h3><p><a href="https://github.com/ianovski/customer-review-sentiment">ianovski/customer-review-sentiment</a></p><h3><strong>References</strong></h3><ul><li><a href="https://podcasts.apple.com/podcast/the-data-skeptic-podcast/id890348705">‎Data Skeptic on Apple Podcasts</a></li><li><a href="https://www.kaggle.com/crowdflower/twitter-airline-sentiment">Twitter US Airline Sentiment</a></li><li><a href="https://stackabuse.com/python-for-nlp-sentiment-analysis-with-scikit-learn/">Python for NLP: Sentiment Analysis with Scikit-Learn</a></li><li><a href="https://code.google.com/archive/p/word2vec/">Code</a></li><li><a href="https://towardsdatascience.com/google-news-and-leo-tolstoy-visualizing-word2vec-word-embeddings-with-t-sne-11558d8bd4d">Google News and Leo Tolstoy: Visualizing Word2Vec Word Embeddings with t-SNE</a></li><li><a href="https://www.kaggle.com/jeffd23/visualizing-word-vectors-with-t-sne">Visualizing Word Vectors with t-SNE</a></li><li><a href="https://www.kaggle.com/varun08/sentiment-analysis-using-word2vec">Sentiment analysis using word2vec</a></li><li><a href="https://scikit-learn.org/stable/auto_examples/model_selection/plot_confusion_matrix.html">Confusion matrix</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de19afeebc08" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Teaching computer science to youth in the Steel City (Hamilton, ON)]]></title>
            <link>https://medium.com/@alexandre3k/volunteering-for-code-org-in-hamilton-on-7d77e1d1708f?source=rss-43fe31977190------2</link>
            <guid isPermaLink="false">https://medium.com/p/7d77e1d1708f</guid>
            <category><![CDATA[education]]></category>
            <category><![CDATA[hour-of-code]]></category>
            <category><![CDATA[hamilton]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Alex Ianovski]]></dc:creator>
            <pubDate>Mon, 28 Nov 2016 04:29:09 GMT</pubDate>
            <atom:updated>2017-01-26T04:36:34.549Z</atom:updated>
            <content:encoded><![CDATA[<p>My own experience with computer science and programming began organically, when I was 15 years old, when the interest first sparked as a result of an introductory high-school course. Since then I’ve studied and trained in different programming languages, both formally and informally, and have been exposed to the exciting world of computer science that continues to push and inspire me on a daily basis. My personal interest in the field, and the fact that it was at such a young age that I first became involved in it, was the reason behind this particular venture.</p><p>Wanting to connect with youth, for the past two months, I have traveled to middle schools across Burlington, Dundas, Hamilton, and Ancaster in order to hold workshops on computer science to students aged 12–14. The purpose of the program is to teach embedded systems programming to middle-school aged kids and get them excited about computer science on the whole. The schools in which I launched these workshops were all unique — for example, the smaller-sized school in Dundas had just 20 students in every grade, compared to a school in Ancaster, similar to the one I attended as a kid, with upwards of 60 students per class. Regardless of classroom size, however, all the schools shared one thing in common — their students were extraordinarily bright! In fact, they were much more adept to the material than I had ever expected.</p><p>Having held similar workshops with university students in the past, I expected that the content of these seminars would be overwhelming for 13 year olds with no previous computer science experience. I expected the students to struggle with the concepts, get distracted, or, at the very least, be intimidated by the tasks. What I experienced was the complete opposite. The determined students were excited about programming and the challenges in front of them. They were persistent, sought out to solve problems, and asked all the right questions when they needed to. I was not only surprised but outright impressed by the level of initiative, critical thinking, and enthusiasm.</p><p>I have learned a number of exciting and valuable lessons from the experiences with these outreach youth workshops. The most pertinent one, personally, seems to be that computer science needs to be integrated as a mandatory requirement, joining traditional science like chemistry and biology, at the middle school level. The students are not only capable or learning the material, they are genuinely enthusiastic about it. As a result, educational systems on both a systemic and individual level should try to find ways to make computer science knowledge not only accessible but integrated seamlessly <em>(pardon the buzz words) </em>into existing curriculum, over a broad range of courses and subject matter. This can be accomplished through formal and traditional classroom delivery, as well as informal and extra-curricular opportunities for students, such as after school programs, student-led clubs and initiatives, or even ventures that bridge learning between parents and students. From my experience with youth engagement I took away many realizations, the most important being that the thirst for knowledge alone is powerful enough to inspire growth and change among youth.</p><p>A link to the workshop material can be found here: <a href="https://goo.gl/etJ40Z">https://goo.gl/etJ40Z</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7d77e1d1708f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>