Author Archives: Yajing Hu

Deblackboxing Algorithms of Recommendation System – Using Two E-commerce Platforms: Amazon and Taobao

  • Abstract

Nowadays the recommendation systems have been widely used in all platforms, especially the e-commerce platforms. Both Amazon in America and Taobao in China have developed their own recommendation system algorithms to find association rules between the customer’s previous consuming behaviors and later purchase and therefore can make a better prediction about that customer’s later purchase. These improved recommendation systems overcome the problems of the previous recommendation system models and can better suit the big data environment.

Keywords: recommendation system, Amazon, Taobao, customer behaviors, big data

 

  • Introduction

The development of network and technology has changed how people communicate. New communication technology has brought many benefits to human beings, such as the faster speed to send and receive information, etc. However, at the same time, it can bring some trouble to people as well. One of the significant disadvantages at this time is information overload. People are surrounded by so much information, and it’s hard for them to find the most relevant one by themselves. Consequently, it’s the latest technology-recommendation system-that helps people again. Recommendation systems aim to overcome the difficulty of finding proper information (Öztürk, & Cicekli, 2011).

Recommendation system provides computer-implemented service that recommends items from a database. It can help customers to find the most relevant data in the large database at a fast speed. Recommendation systems make use of techniques such as information retrieval, customer modeling, and machine learning. The recommendations are customized to particular customers based on the data and information known about the customers. In the recommendation system, the input includes the attributes of the products and the attributes of the customers, while the output can reflect as a numeric score that is a measure of how much the algorithm “believes” that a particular customer will enjoy the recommended content or buy the recommended products.

Recommendation systems work in different domains. For example, the Netflix recommendation system aiming to suggest relative matches to its customers provides a various number of movies and TV shows. YouTube recommendation system can “guess” users’ preferences based on data and put what users might be most interested in from millions of videos on each users’ landing page. One other common application for recommendation system involves recommending products to online customers on the e-commerce platforms. E-commerce platforms believe that people do not buy things at random. Their purchases depend on a number of factors, and if the recommendation system algorithms can estimate such factors for each customer, it can make an accurate prediction about that customer’s later purchase (Jacobi, Benson, & Linden, 2011). Recommendation system can also find association rules with the form “People who buy X are also likely to buy Y”, which can analyze customer behaviors and therefore make the right products reach their potential customers.

 

  • Three Main Types of Recommendation System

Recommendation systems can be broadly divided into three main types according to the approaches they use to make recommendations, namely content-based recommendation, collaborative recommendation, and hybrid recommendation (Öztürk, & Cicekli, 2011).

  • Content-based recommendation

Content-based recommendation system can also be called feature-based recommendation system. It analyzes the features of the content in the set and matches them to features of the customers, based on a user model developed by analyzing the previous action of the customers (Wang, Zhang, & Vassileva, 2010). Pure content-based recommendation systems suggest items based on their similarity to the content of the items which the customers selected before. For social media recommendation such as Instagram and Facebook, the content-based recommendation is quite useful because they can find similarity in the recommended content.

However, content-based recommendation systems generally do not provide any mechanism for evaluating the quality or popularity of an item (Jacobi, Benson, & Linden, 2011). In addition, one significant prerequisite of content-based recommendation systems is that the items must include some form of content that can feature extraction algorithms. As a result, content-based recommendation systems can only be limited to social media. It is not suitable for recommending products, movies, books, restaurants, and other types of items which include little or no useful content to be distinguished by algorithms.

  • Collaborative filtering

Collaborative filtering recommendation system can also be called the social recommendation system. In the collaborative filtering recommendation system, items are recommended to customers based on the interests of a community of customers, without the analysis of items’ content (Jacobi, Benson, & Linden, 2011). Items are suggested according to the similarity between users with similar habits. Collaborative filtering recommendation system relies on customers to rate individual items from a list of popular items. Through the process, each customer builds their own personal profile of rating data. By statistically correlating users based on their previous profile and based on the assumption that people who behave similarly in the past will continue to buy similar products in the future, items that were rated highly by one group customer will be recommended to similar customers in their “community”. One important benefit of the collaborative filtering recommendation system is that it overcomes the previous deficiencies of content-based recommendation system and it is widely used to recommend movies, books, or products on e-commerce platforms.

Of course, pure collaborative filtering recommendation systems have deficiencies as well. One significant problem is that customers might feel tired to rate items in the database to build up his or her own personal profile. This process can be frustrating and time-consuming, particular in case customers are not familiar with many of the items that are presented for rating. In addition, since rating process is the prerequisite for this recommendation system, if an item is just launched on the e-commerce platforms and has not been rated by many customers yet, it might face with a “cold start” problem in which the service cannot be brought online until a threshold quantity of rating data has been collected (Jacobi, Benson, & Linden, 2011). Furthermore, since the collaborative filtering recommendation system relies on the community of “similar customers”, it is poorly suitable for providing recommendations to customers who have unusual tastes.

In addition, both content-based recommendation system and collaborative filtering recommendation have one common disadvantage. Generally, they can not reflect the latest preferences of the customers. Both the recommendation systems need time to update the data so that the recommended results can be refreshed. This might result in the inaccurate recommendation outcomes and even recommend items that customers used to purchase.

  • Hybrid recommendation system

Hybrid recommendation system combines the previous methods to obtain better performance. For example, Adsorption is a hybrid recommendation system that aims to select appropriate videos or movies for customers applied to YouTube successfully. In YouTube, there are millions of videos available, and customers can customize their own feeds based on big data and machine learning. First, rating information is collected according to the collaborative filtering recommendation system. And then, content-based features are injected to provide a hybrid system. Adsorption uses this rating information and tries to reach unrated videos using a graph-based algorithm.

Fig. 1. General System Architecture of YouTube Recommendation System

Retrieved from: https://link.springer.com/chapter/10.1007/978-3-642-21827-9_42

 

  • Advantages of Recommendation System Used in E-Commerce Platforms

Recommendation system can save customers’ time and energy in finding their needed information online. Processing information in the web pages and navigation on the web can take a significant amount of time for customers, requiring them to employ higher cognitive processes such as generalization and categorization (Ševce, Tvarožek, & Bieliková, 2010).

Customer experience personalization is also all about data. By applying big data and machine learning, recommendation systems can shape the overall customer experience. As Steve Jobs said, “At a lot of times, people don’t know what they want until you show it to them.” Recommendation system can provide a great way to help people “discover” items that they will like but unlikely to discover by themselves because they might easily get lost in large amounts of data. Recommendation system can improve a visitor’s experience by offering relevant items at the right time and on the right page. How well recommendation systems boost subscriber numbers through engagement and stickiness, facilitating such serendipitous discovery has turned into a high stakes multi-billion-dollar race for the world’s biggest digital companies (Arora, 2016).

A good recommendation system is able to react immediately to changes in a customer’s data and makes a compelling recommendation for all customers regardless of the number of purchases and ratings with fast speed, and the processing normally only requires sub-seconds.

 

  • Previous E-commerce Recommendation Algorithms and Their Disadvantages

The recommendation system is best known for its use on e-commerce platforms, where they input a customer’s interests and attributes to predict their potential consuming behaviors and generate a list of recommended items. The attributes of customers include not only their purchase and ratings, but also the items they viewed, the demographic information of each customer, and their subject interests, which can all be analyzed as the input data by the recommendation system algorithms.

However, e-commerce recommendation system algorithms face some challenges. First, although the data can be processed at a fast speed, customer data is volatile, and each interaction between customers and products are valuable so that it’s hard for the algorithms to respond to all the new information immediately. Second, the speed of reaction is important, and the quality is also important. Many applications require the recommendation results to be returned in real time and at a high quality. Third, for new customers, they typically have extremely limited information, based on only a few purchases of product ratings. Consequently, the recommendation results based on such little data can be not as accurate as older customers. While last but not least, for older customers, the recommendation system still faces the challenge that they have a glut of information based on thousands of purchases and ratings, and some contradictory information, for example, might also influence the outcomes of the recommendation systems.

Most previous recommendation systems applied on e-commerce platforms by looking for customers who purchased and rated similar items and grouped these customers into a community (Linden, & Smith, 2003). For each customer in the community, the algorithm aggregates the items, eliminates the items which have been bought before, and recommends the remaining items with high ratings to the customer. This belongs to the collaborative filtering recommendation system, which focuses on similar customers, rather than the content of the items.

One other kind of recommendation system algorithm, namely the content-based recommendation system, focuses on similar items rather than the similar customers instead. For each of the user’s purchased and rated items, the algorithm attempts to find similar items. It then aggregates similar items and recommends them (Linden, & Smith, 2003).

  • Traditional Collaborative Filtering

The goal of collaborative filtering algorithm is to suggest new items or to predict the utility of a certain item for a particular customer based on the customer’s previous liking and the opinions of other like-minded customers (Sarwar, Karypis, & Konstan, 2001). A traditional collaborative filtering algorithm can represent a customer as an N-dimensional vector of items, where N is the number of distinct catalog items. The traditional collaborative filtering algorithm can rank each item according to how many similar customers have purchased or rated it and select recommendations from similar customers’ items using various methods to measure similarity. One common method is to measure the cosine of the angle between the two vectors, using the formula as follows.

Formula to measure the similarity of two customers, A and B

Retrieved from: https://www.computer.org/csdl/magazine/ic/2003/01/w1076/13rRUwwaKpC

However, the vector is quite sparse for very large data sets. Therefore, to generate high-quality recommendation outcomes with the traditional collaborative filtering algorithm is expansive. The high cost might be the leading cause that most e-commerce platforms will not use this algorithm to recommend items to their customers.

  • Cluster Models

The cluster model algorithm divides the customer base into many segments according to some specific standards to group the most similar customers and turns the recommendation process into a classification problem. The goal of the cluster model algorithm is to assign the user to the segment containing the most similar customers.

Unlike the traditional collaborative filtering algorithm, cluster models can solve the sparse problem because it compares each user to a controlled number of segments rather than the entire customer base (Linden, & Smith, 2003). However, the recommendation quality is low at the same time because the standard of classification is relative. The cluster models might not find the right group that includes the most similar customers, and therefore the recommendations it produces are less relevant.

  • Search-based Methods

The search-based method is one kind of content-based method, which treats the recommendations problem as a search for related items. Given the user’s purchased and rated items, the algorithm constructs a search query to find other popular items by the same author, artist, or director, or with similar keywords or subjects (Linden, & Smith, 2003).

With the collected and created profiles, the search-based methods can alleviate the cold-start problem of collaborative filtering algorithms. However, this search-based method algorithm also has disadvantages, too. The recommendations are often either too general or too narrow. Besides, the search-based method performs better in terms of content and its significant attributes, which limits the use of such algorithms.

 

  • Case 1: Amazon and Item-to-item Collaborative Filtering

Unlike the previous recommendation system algorithms, in this paper, I will use the two special recommendation systems used in Amazon, the largest e-commerce marketplace and cloud computing platform in America, and Taobao, the biggest e-commerce website in China, as two main cases and analyze and compare the two different recommendation systems.

Amazon, as the largest e-commerce platforms in the world, is now investing a large amount of talent and resources to integrate artificial intelligence, especially “deep learning” technology to make the recommendation system more efficiently. Amazon’s recommendation system is called item-to-item collaborative filtering. Amazon uses this algorithm to integrate recommendations across the buying experience, from product discovery to checkout and personalize the online store for each customer. The effective pinpoint of predicted products to specific customers according to their profiles by recommendation systems can vastly exceed those of untargeted content such as banner advertisements and top-seller lists (Linden, & Smith, 2003).

In Amazon’s item-to-item collaborative filtering algorithm, the input data includes customers’ purchase history, items in their shopping cart, items they have rated and liked, and what other similar customers have viewed and purchased (Arora, 2016). Item-to-item collaborative filtering algorithm matches the customer’s purchased and rated to similar items and then combines those similar items into a recommendation list.

To ensure the quality of the outcome of the item-to-item collaborative filtering algorithm and provide the most similar match for each given item, the algorithm builds a similar-items table with a product-to-product matrix to find items that customers tend to purchase together. The product-to-product matrix is used to iterate through all item pairs and computing a similarity metric for each pair (Linden, & Smith, 2003). The degree of similarity between two items does not totally depend on common customers. The following iterative algorithm provides a better approach by calculating the similarity between a single product and all related products.

For each item in product catalog, I1

                        For each customer C who purchased I1

For each item I2 purchased by customer C

                                                Record that a customer purchased I1 and I2

For each item I2

                        Compute the similarity between I1 and I2

Iterative algorithm to calculate the similarity between a single product and all related products

Retrieved from: https://www.computer.org/csdl/magazine/ic/2003/01/w1076/13rRUwwaKpC

Given a similar-item table, the algorithm finds items similar to each of the customer’s purchases and ratings, aggregates those items, and then recommends the most popular or correlated items (Linden, & Smith, 2003).

Amazon not only used big data in the recommendation system but also opened up its sophisticated artificial intelligence technology as a cloud platform. In May 2016, Amazon unveiled its DSSTNE, which is an open source artificial intelligence framework that Amazon developed to power its own product recommendation system (Arora, 2016). The DSSTNE can be used as an open source software so that the promise of deep learning can extend beyond speech and object recognition to other areas such as research and recommendations. It can lead to better prediction based on fewer data and therefore customers are more likely to click on and buy those products recommended by the recommendation systems.

 

  • Case 2: Taobao and Tree-based Deep Model

Taobao is a Chinese online shopping website owned by Alibaba. It’s the world’s biggest e-commerce website and the seventh most visited website according to Alexa. Recommendation, search, and advertisements placement are all core tasks for providing internet content and data distribution to e-commerce businesses. The task of dealing with data in Taobao’s recommendation system is hard because the amount of data is huge. Alibaba collects data related to everything from a customer’s purchase history to the pages they view to products they bookmark. Taobao divides its customers into 500 different segments and couples those segments with the information it has on more than 1 billion products being sold on the website before putting artificial intelligence to work to generate the most accurate recommendations possible (Zhu et al, 2018).

Taobao’s special recommendation system algorithm is called Tree-based deep recommendation model, abbreviated for TDM. The architecture of TDM can be reflected in the following figure.

Fig. 2. The system architecture of Taobao display recommendation system

Retrieved from: https://dl.acm.org/citation.cfm?id=3219826

The recommendation algorithm process can be broken into two phases, Matching and Ranking. The first step is matching. First, after receiving page view request from a customer, the system uses user features, context features, and item features as input to generate a relatively smaller set of candidate items from the entire corpus in the matching server. The corpus is huge and always includes hundreds of millions of items, while the number of the filtered set of candidate items is usually hundreds (Zhu et al, 2018).

The second step is ranking. With hundreds of candidate items, the real-time server uses more expressive but also more time-consuming models to predict indicators like click-through rate or conversion rate. After ranking by strategy, several items are ultimately impressed to each customer.

The reason why Taobao adopted tree-based deep recommendation model is that it can solve the difficulty of information overload. E-commerce platforms typically have a gigantic corpus, so that the cost of using algorithms to predict each customer’s preference is huge and therefore the process of full corpus retrieval rather problematic. But with TDM algorithm, the huge data in the corpus is filtered into a smaller set at the beginning of the recommendation process, and this algorithm does not need to meet the challenge of huge amounts of data. In addition, recommended items also have to be as novel as possible. The interaction between customers and products is valuable and might be updated all the time so that even little change might influence the outcome of the recommendation system. Results that simply replicate a customer’s previous behavior are undesirable. In this respect, memory-based and item-based collaborative filtering both fall short (Zhu et al, 2018).

As a comparison, the tree-based deep recommendation model can make a huge amount of information manageable. Frustrated with the shortcomings of existing models, the Alibaba tech team decided to develop this novel tree-based deep recommendation model, which can leverage a hierarchy of information and turns the recommendation problems into a series of hierarchical classification problems (Zhu et al, 2018). With this tree-based deep recommendation model, more accurate and efficient prediction can be made from a large corpus. Like Amazon, apart from the recommendation system, Alibaba’s cloud computing subsidiary also developed an operating system called Apsara that organizes data centers into a computational engine that can process more than 175,000 transactions per second, which can also embody its efficiency with its newest high technology to better serve customers.

 

  • Conclusion

The recommendation system is widely used in the e-commerce platforms to put the most potential products on each customer’s landing page and help customers find the most relevant products they want. However, previous existing algorithm models cannot keep pace with new technology and therefore new algorithm models are required. Both Amazon and Taobao use their special recommendation systems, namely item-to-item collaborative filtering and tree-based deep model. These two recommendation systems have some common characteristics. Both of them are integrated with cloud computing to provide better service to their customers. Their recommendation processes both need to filter the data first to improve the efficiency, and then the ratings are ranked to make the outcomes of the recommendation system with higher quality and lead to better and more precise prediction. At the same time, these two recommendation systems have some differences. In the recommendation algorithm, Amazon uses the item-to-item matrix, while Taobao uses user-candidates matching.

 

References

  1. Arora, S. (2016). Recommendation engines: How Amazon and Netflix are winning the personalization battle. MarTech Advisor.
  2. Errico, J. H., Sezan, M. I., Borden, G. R., Feather, G. A., & Grover, M. G. (2015). Collaborative Recommendation System. U.S. Patent No. 8,949,899. Washington, DC: U.S. Patent and Trademark Office.
  3. Li, B., Qian, C., Li, J., Tang, K., & Yao, X. (2016, July). Search based recommender system using many-objective evolutionary algorithm. In 2016 IEEE Congress on Evolutionary Computation (CEC) (pp. 120-126). IEEE.
  4. Linden, G., Smith, B., & York, J. (2003). Amazon. com recommendations: Item-to-item collaborative filtering. IEEE Internet computing, (1), 76-80.
  5. Öztürk, G., & Cicekli, N. K. (2011, June). A hybrid video recommendation system using a graph-based algorithm. In International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems (pp. 406-415). Springer, Berlin, Heidelberg.
  6. Pazzani, M. J., & Billsus, D. (2007). Content-based recommendation systems. In The adaptive web (pp. 325-341). Springer, Berlin, Heidelberg.
  7. Sarwar, B. M., Karypis, G., Konstan, J. A., & Riedl, J. (2001). Item-based collaborative filtering recommendation algorithms. Www, 1, 285-295.
  8. Ševce, O., Tvarožek, J., & Bieliková, M. (2010, September). Term ranking and categorization for ad-hoc navigation. In International Conference on Artificial Intelligence: Methodology, Systems, and Applications (pp. 71-80). Springer, Berlin, Heidelberg.
  9. Wang, Y., Zhang, J., & Vassileva, J. (2010, September). Towards effective recommendation of social data across social networking sites. In International Conference on Artificial Intelligence: Methodology, Systems, and Applications (pp. 61-70). Springer, Berlin, Heidelberg.
  10. Zhu, H., Li, X., Zhang, P., Li, G., He, J., Li, H., & Gai, K. (2018, July). Learning Tree-based Deep Model for Recommender Systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (pp. 1079-1088). ACM.

Deep Learning’s Pros and Cons

Deep learning is essentially a statistical technique for classifying patterns, based on sample data, using neural networks with multiple layers. Deep learning is a subset of machine learning. Deep artificial neural networks are a set of algorithms that have set new records in accuracy for many important problems, such as facial recognition, sound recognition, recommendation systems, natural language processing, etc. Although it has brought much convenience for human beings, people still have a critical appraisal against it. I want to take the facial recognition system as an example.

The facial recognition system is a technology capable of identifying person from a digital image or video. It works by comparing selected facial features from given image with faces within a database. Consequently, in order to make sure that the outcome is accurate, the deep learning works best when thousands, millions, or even billions of examples or data have been collected in the database. However, if the database is not big enough, the validity of the outcome will no doubt be questioned.

In addition, the safety of information can also be doubted in the facial recognition system. If your face can directly lock your phone, your other personal information will become visible to others as well. And others might also use your face to log in your bank account and steal your money. The source of the facial information can easily retrieve from the social media, from any selfies you posted on the internet. You don’t hope that your facial password is always exposed to anyone else, right?

Also, the feasibility of the facial recognition system can be doubted. Deep learning presumes a large stable world, while in real world, it can be constantly changing. If the system makes some mistake, while people trust this system, the outcome might be serious. The source of the database for the facial recognition system is difficult as well. The facial information is hard to be captured online. If the facial information is captured offline by police and used online, when the database was hacked, or destroyed, the outcome can be so dangerous and horrible.

In the end, this technology is also related to moral issues, and that is why mosaic was used in the news and other videos. Face, as one of the people’s personal information, has the right of privacy like others. The facial recognition system no doubt violates this right and will raise moral debate among people, although it is quite interesting and convenient.

 

Questions/Topics:

I am quite interested in the information safety of the cloud system, this is more significant in the public cloud and community cloud. Since the information and resource can be shared in cloud through the internet, does it mean that information will not be private? What if a company wants to enjoy the benefits of the cloud system as well as keep the information secret to a certain group of people without being violated by others?

 

Reference:

  • Agre, P. (1997). Computation and human experience. Cambridge; Cambridge University Press.
  • Thorn, P. (2015). Nick Bostrom: Superintelligence: Paths, Dangers, Strategies. Minds and Machines25(3), 285–289. https://doi.org/10.1007/s11023-015-9377-7

Big Data & Advertising

The purpose of advertising and marketing is to inform, educate, and ultimately persuade consumers to buy a given product or service. According to the STP strategy, namely Segment, Target, Position in marketing, defining the target audience is very important and advertisers need to analyze data to resonate with their target audience. Different types of audience will prefer and resonate with different content. Advertisers will first need to group different people by age, economy, or other factors and then use different market strategy to respond to each group.

These are all about big data. Big data enables companies to better target the core needs of customers by developing rich and informative content. Big data can be considered as the huge amount of information which are available to anyone in the world over the internet. Big data can help advertisers gain essential insights into their target demographics, such as patterns, consumer habits and trends, etc. Using artificial intelligence to extract and analyze data from many avenues such as subscriptions, followers, browser history, search record can understand the audience’s preference better. These figures generate insights that can lead to better business decisions and strategic moves. The application of the right technology improves the quality of decision making and detailing processes. When the transaction has been changed from offline to online, the data has been more digitalized so that the artificial intelligence technology to analyze such huge amount of data.


https://www.mentionlytics.com/blog/5-real-world-examples-of-how-brands-are-using-big-data-analytics/

This is similar to the recommendation system. For example, Netflix uses big data analysis for target advertising. In this case, the big data is from over 100 million subscribers, their past search and view history. The use of artificial intelligence and big data can help uncover the hidden patterns, correlations and give insights so as to make proper business decisions and customize audience’ preference. The modern data analytics systems allow for speedy and efficient analytical procedures. This ability to work faster and achieve agility offers a competitive advantage to businesses.

Reference

  • Bridgewater, R. (2016). O’Neil, Cathy. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. (Brief article) (Book review). Library Journal, 141(19).
  •  Kitchin, R. (2014). Big Data, new epistemologies and paradigm shifts. Big Data & Society, 1(1). https://doi.org/10.1177/2053951714528481

Combining AI / ML with Cloud Computing

Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computer resources that can be rapidly provisioned and released with minimal management effort or service provider interaction (Ruparelia, 2016). Nowadays, nearly every application we use today is paired with a cloud backup service and it’s cheaper and more convenient for data access and storage. Cloud computing enables end-users to access to real-time and on-demand self-service from anywhere in the world. Additionally, end-users only need to pay for the service they consume as they consume it so that the cost is cheaper than traditional backup methods.

With the combination of artificial intelligence and cloud computing, the end-users can basically get amounts of data which are stored through the virtualized network. Additionally, the algorithms enable the capacity to learn and improved on the go. For example, IBM’s question-answering computer system Watson, gives end-users the freedom to bring its services to all of their data on any cloud platform. One of the problems faced by organizations before is that their data is spread across a variety of environments such as public clouds, private clouds, etc. However, this year, IBM announced a new way to make Watson available anywhere, no matter on private cloud, public cloud, community cloud or hybrid cloud. They built and integrated a series of new Watson micro-services on IBM Cloud Private for Data, which is IBM’s open, cloud-native information architecture for artificial intelligence. This cloud, combined with data systems, helps companies uncover previously unobtainable insights from their data, wherever they reside.

The other example of combing machine learning and cloud computing is Google’s Cloud ML Engine.

ML workflow, retrieved from: https://cloud.google.com/ml-engine/docs/tensorflow/technical-overview

Google’s Cloud Machine Learning Engine is a managed service that enables developers and data scientists to build and bring superior machine learning models to production. With the advantage of machine learning and data analysis, Google’s Cloud ML Engine can find patterns in data, make accurate determinations or predictions via data in the cloud. For the training of the computer model, Cloud ML Engine enables users to automatically design and evaluate model architectures to achieve an intelligent solution faster and without experts. As a cloud system, it has rapid elasticity and it can scale to leverage all the data. Once the users have a trained model, prediction applies what the computer learned to new examples.

However, there are some disadvantages remaining to be solved for the future combination of cloud computing and artificial intelligence. The security of ownership and privacy of data and compliance issues are still the problems. Additionally, the lack of understanding and knowledge still exists, no matter of artificial intelligence or cloud computing, which might cause fear for those companies.

 

References:

Cristopher, B. (2018). Retrieved from: https://bigdata-madesimple.com/pros-and-cons-of-combining-artificial-intelligence-and-cloud-computing/

Rob, T. (2019). Retrieved from: https://www.ibm.com/blogs/think/2019/02/enabling-watson-anywhere/

Rountree, D., & Castrillo, I. (2013). The basics of cloud computing: Understanding the fundamentals of cloud computing in theory and practice. Newnes.

Ruparelia, N. B., & Ruparelia, N. (2016). Cloud computing. Mit Press.

Will, K. (2016). Retrieved from: https://www.technologyreview.com/s/602744/ibms-watson-is-everywhere-but-what-is-it/

Face Recognition and Ethical Issues

Face recognition is a method of identifying or verifying the identity of an individual by using their face due to artificial intelligence. Face recognition is playing an increasing role in law enforcement, border security and other purposes around the world. Although it has many advantages and benefits, it still raises some ethical concerns.

One of the important ethical issues is about privacy. Once someone has your face print, they might get your name, your age, your birthday date, and then they can find your social account, and even track you in the street, track your home location, etc. They might even unlock your phone and bank account. For example, Alipay, China’s leading online paid solution, announced a major upgrade to its “Smile to Pay” service in 2018, aiming to make its face recognition process more accessible to merchants and customers. This facial recognition payment seems to be very convenient and safe, but if someone post on his/her selfie on the social account that anyone can see and save, others might invade his/her Alipay account and pay with the recognizable face, without needing to input the password.

Besides, another ethical issue is about surveillance in the face recognition technology. China also uses face recognition to find traffic violators and “shame” jaywalkers. In Shanghai, people always find that they receive the message to inform that they ran red light at certain time at certain street. This is because the camera captures their face and recognize it from the database. On one hand, this artificial intelligence technology can help regulate public order. On the other hand, this might make people have the feeling of living under the surveillance, even in the public sphere.

 

Reference

Siri and Dialects

With the advanced deep learning functionalities of automatic speech recognition for converting speech to text and natural language understanding to recognize the intent of context, voice-assistant software has been released, such as Siri, Amazon Lex, Google Assistant, etc. Among these, Siri is the oldest and although it has squandered its lead when it comes to understanding speech and answering questions, it still has the advantage of speaking 41 languages (ios 12.1 version), which is a very important capability for a smartphone in the global market. For example, it has 9 different kinds of English in Siri, including Australia, Canada, India, Ireland, New Zealand, Singapore, South Africa, United Kingdom, and United States. For Chinese, it is not only divided into Mandarin and Cantonese, but each can also target at different markets.

In recent years, Apple started to work on new languages by bringing in humans to read passages in a range of accents and dialects, which requires Apple to capture a range of sounds in a variety of voices. However, people with strong accents who use regional words are having to change how they speak almost entirely in order to be understood by Apple’s Siri. For example, when I speak Nanjingese, which is a kind of dialect used in Nanjing, China, Siri can’t understand what I say and ask me what I am talking about.

Speaking languages natively is complicated for any assistants. Although it is a painstaking process, the language barrier is important to overcome for voice assistants if they are to become ubiquitous tools for consumers globally. Digital assistants should use more colloquial language in the future and Siri has always been updating itself in order to understand more languages and dialects. For example, in March 2017, Apple reports that soon Siri will be able to speak Shanghainese, a dialect of Chinese spoken only around Shanghai, since China alone makes up almost 33% of the smartphone market. However, it is now 2019, but Siri still cannot speak Shanghainese. But the keyboards can add Shanghainese as one of the dictation languages and it still belongs to the speech recognition application.

And I tried this (Although I can only speak ‘Thank you’ in Shanghainese), it really works!

Apple not only captures a range of sounds to predict words sequences in order to understand accents, but also deploys “dictation mode”. When customers use dictation mode, Apple captures a small percentage of audio recordings. After enough data has been gathered, Siri is released with answers to what Apple estimates will be the most common questions. Once released, Siri learns more about what real-world users ask and is updated every two weeks with more tweaks.

Maybe in the future, we can expect to Siri to speak dialects in the synthesized voice.

Reference

Ava, M. (2017). Apple’s Siri to Soon Speak Shanghainese, 21 Languages in Total. Retrieved from: https://voicebot.ai/2017/03/09/apples-siri-soon-speak-shanghainese-21-languages-total/

Sarah, G. (2016). Is Siri killing off regional dialects? Users admit to faking accents to make themselves understood by the voice assistant. Retrieved from: https://www.dailymail.co.uk/sciencetech/article-3442173/Will-Siri-kill-regional-accents-Users-admit-faking-dialects-make-understood.html

Stephen, N. (2017). Apple’s Siri learns Shanghainese as voice assistants race to cover languages. Retrieved from: https://www.reuters.com/article/us-apple-siri-idUSKBN16G0H3

Difficulty in Machine Translation, From English to Korean

The latest two language translation technology models based on statistics or artificial neural machine translation, namely Statistical Machine Translation and Neural Machine Translation. Statistical Machine Translation requires an enormous amount of data while Neural Machine Translation utilizes a large neural network and the deep learning which make it possible to acquire context-sensitive translation.

Korean, as well as Chinese and Japanese, belong to the Han Ideographs. English and Korean are significantly different in terms of structure, such as the distribution of subjects, the word order, the forms of verbs, and so on. English uses a Subject-Verb-Object structure, while Korean uses a Subject-Object-Verb structure. For example,

Besides, omitting subjects in Korean creates confusion in comprehending the meaning. For example,

As you can see, the ‘영희가’ can be omitted in Korean, which can cause a problem in understanding the sentence. Consequently, in order to avoid this, the context in a discourse needs to be closely considered, and this requirement works as a challenge for Machine Translation.

Another difficulty is that in Korean, speech is divided into polite form and impolite form, depending on who you talk to, which is extremely important in Korean since if used inappropriately, it seems quite rude. And the differences between polite form and impolite form is complicated. The politest and formalist form of speech is ending in ‘십니다’, while the less polite and formal form is ending in ‘~요’. And the above two polite speech is when talking to the elder, superiors, or people that you are not familiar with. For example, ‘I listen to music’ in Korean is ‘음악을듣습니다’ or ‘음악을들어요’ when speaking in a polite form. However, when you talk to your friend, subordinate, or people younger than you, the same English sentence will be translated to ‘음악을듣다’ instead, which is an informal and impolite form of speech in Korean. The application of polite or impolite is totally dependent on the context, sometimes a younger person can still talk to an older person in an impolite form, if they are close friends or the older person agree with this.

Besides, the usage of 1stperson and 2ndperson is different in polite and impolite form of speech in Korean. For example,

when I input ‘Do you eat lunch’ in Google translate, the translated one is quite impolite, and it’s very rude to ask like this to the elder, etc. If you watch some Korean dramas, you might find the first character, ‘너’(neo), which means ‘you’ is common to be seen when you look down on the others, and it is quite rude. The ‘I’ and ‘You’ are different in polite and impolite form of speech in Korean. ‘I’ is’저’ in polite form and ‘나’ in impolite form, while “You” is ‘당신’ in polite form and ‘너’ in impolite form.

Also, there are some differences depending on the gender of people. In Korean, people seldom directly call the name of the elder, even if their ages gaps are small. For example, a 13-year-old girl must call a 14-year-old ‘sister’, or else it will be rude. The word ‘older sister’ and ‘older brother’ is used by both boys and girls in English, however, Korean use different words depending on their gender. If a girl call her older sister, she must say ‘엄니’. If a boy calls his older sister, he must say ‘누나’. If a girl calls her older brother, she must say ‘오빠’. If a boy calls his older brother, he must say ‘형’.

These differences significantly add to the complexity and difficulty in machine translation from English to Korean. And here is a video talking about some problems with machine translation from English to Korean and why Korean Machine Translation is terrible.

 

Reference:

Teller, V. (2000). Speech and language processing: An introduction to natural language processing, computational linguistics, and speech recognition. Computational Linguistics26(4), 638-641.

Kim, S., & Lee, H. (2017). A Study on Machine Translation Outputs: Korean to English Translation of Embedded Sentences. 영어영문학, 22(4), 123–147.

Data Science and Amazon Recommendation

Data science focuses on improving decision making through the analysis of data. After collecting many various kinds of data in large amount, different types of patterns can be analyzed and extracted, which helps us identify groups of customers exhibiting similar behaviors and tastes, which helps customer segmentation in business. AI helps when we had a large number of data examples and when data patterns are too complex for humans to discover and extract manually (Kelleher & Tierney, 2018). One important use of data science is in sales and marketing, namely recommendation system.

Judging by Amazon’s success in market, artificial intelligence is increasingly playing an important role in Amazon’s competitive advantage. And in which, two of the best applications of artificial intelligence are including on-site and off-site product recommendations.

Amazon’s recommendation system is based on unsupervised learning, and its aim is to find the regularities and patterns in the input to see what normally happens, therefore find clusters or groups of input which has structure embedded in the data (Alpaydin, 2016). Normally, people do not buy things randomly. Instead, there are certain association rules inside the behaviors and their purchase depend on a number of factors. For example, demographic information. Amazon recommendation system algorithm takes these data as input and group these input with artificial intelligence to make prediction about the existing customers’ later purchase and attract potential customers at the same time. Besides, there are certain hidden factors and if AI can estimate those hidden factors for a customer, it can make more accurate estimation. This is all about data and the invisible pattern between input and output.

For the data mining of Amazon recommendation system, we first need to figure out its input data and data source, which includes purchased shopping cart, items added to carts but abandoned, wish lists, dwell time, referral sites, customers’ demographic information, number of times viewed an item before final purchase, click paths in session, pricing experiments online, etc. These data are so huge, but with artificial intelligence, it can easily find the hidden factors and invisible patterns and generate the “Recommended for You, XXX” section on the website which leads customers to a page full of products recommended just for each individual customer. That is to say, artificial intelligence can create a personalized shopping experience for every customer.

Amazon’s recommendation system can also generate a “Frequently Bought Together” section which is found below every product listing and suggests a combination of complementary products. The focus here is on cross-selling products to increase order size. And this section is quite important because people might need the complementary products rather than contradictory ones, so that the choice of products recommended in this section should be cautious and it’s better to recommend a group of products that customers can buy as a bundle.

Besides, Amazon has a section named “Customers Who Bought This Item Also Bought” section, which is similar to “Frequently Bought Together”. In this section, Amazon display items which have been purchased together in the past to increase the average order values through cross-selling.

Amazon also looks at the products which customers have been browsing and thinks of the reasons why customers see it but not buy it with data. Besides, Amazon might guess customer’s psychological factors and recommend them very similar products of different shapes, sizes, and brands to help them find products which customers might be interested. Also, this process act like visiting the physical stores off-line, where customers can compare the same products of different brands and make wiser purchase decision after comparing immediately.

Reference:

Alpaydin, E. (2016). Machine learning: the new AI. MIT Press.

Boden, M. A. (2016). AI: Its nature and future. Oxford University Press.

Justin, Y. (2017). 5 lessons you can learn from Amazon’s recommendation engine. Retrieved from http://altitudelabs.com/blog/amazon-product-recommendation-engine/

Information Theory and JPEG Compression

The simplest kind of information system is a communication system, and the communication system includes encode, decode, transport, store, and retrieve data or signals. All formats of data, which includes numbers, signals, logic formulas, and text, etc. can be represented as patterns of bits. According to the information transmission model by Shannon, the starting point of transmission process is source, and the destination is the end of the tour. During information transmission, the message is encoded and readable by transmitter in the form of signal, and after being influenced and altered by noise in the environment, the signal as well as the noise can be received by the receiver. Then after decoding, the information can finally reach the destination.

The basic flow of image compression coding

One of the most important application of Shannon’s information theory is image compression. The main goal of image compression is to store and transmit data in an efficient form. Since the transmission chases for efficiency, the quality of image can be lossy, or lossless. JPEG is a commonly used method of lossy compression for digital images. It stands for Joint photographic experts group. It is the first international standard in image impression and widely used till today.

In the start of this process, the representation of the colors in the image is converted from RGB to YUV. Y represents brightness of a pixel, U and V which are two chroma components represents color, including tints and shades, and saturability. Since human eyes are more sensitive to the difference of brightness than that of color, so that in this process, the resolution of the chroma components is reduced. And this process is called downsampling. Then, each channel is split into 8*8 blocks, and the range of the pixels intensities now are from 0 to 255. Then we need to subtract 128 from each pixel value from -128 to 127. The next step is to use DCT, which stands for discrete cosine transform and round to the nearest integer. Because human eyes are not good at distinguishing the high frequency brightness variation, it allows us to reduce the amount of information in the high frequency components. Also, since many higher frequency components are rounded to zero and many others become small positive or negative numbers, the pixel takes fewer bits to represent.

the original image (left) and compression image (right)

The DCT-based image compression such as JPEG performs very well at moderating bit rates; however, the quality of the image decreases because of the encoding and decoding process so that the resolution of the transmitted image might not be as high as the previous one. With its important advantages, JPEG still becomes the most common format for storing and transmitting photographic images on the World Wide Web.

 

Reference:

Candès, E. J., & Wakin, M. B. (2008). An introduction to compressive sampling. IEEE signal processing magazine25(2), 21-30.

Denning, P. J., & Martell, C. H. (2015). Great principles of computing. MIT Press.

Floridi, L. (2010). Information: A very short introduction. OUP Oxford.

Wei, W. Y. (2008). An introduction to image compression. National Taiwan University, Taipei, Taiwan, ROC.

 

The Mystery Behind Instagram Recommendation System

Machine learning is a kind of new data-centric approach which relies on data to enable artificial intelligence capable of behavior as intelligent exhibited by humans. Ranking is an application area of machine learning which train on pairs of instances and have the outputs for the two in the correct order (Alpaydin, 2016, P81). Instagram is now adding one recommendation system which suggests posts for the audience based on those that have been liked by other accounts they follow. The recommendation system relies on machine learning based on each audience’s past behavior and its aim is to create a personalized feed based on how they interact with other accounts.

The new section, “Recommended for You”, although is similar to advertisements, is clearly labeled so as not to be confused with audience’s own home feed. In which, it contains the suggested posts which was recommended according to the algorithm. Three main factors determine what audience might see in their Instagram feed, namely interest, recency and relationship. Beyond these core factors, there are other factors which more or less influence the ranking system, such as frequency, following, and usage habit.

Like Apple’s Siri, Instagram Recommendation System is a rule-based personal assistant which can focus on broadening users’ access to content and switch from chronological feed to algorithmic one based on both topical relevance and personal relevance. Such relevance is related to weighted search and data mining, which are two of the prominent natural language processing applications of information retrieval (Boden, 2016, P63). Certain posts are assessed statistically and weighted by relevance, while data mining can find patterns unsuspected by human users. After data analysis, such posts are ranked according to their weight and relevance, which can be reflected by statistics and then they can be recommended to each user according to their ranks.

In the Instagram recommendation system, the algorithm collects all the data of the example observations and analyzes it to discover the relationship which might not be observed by humans. The input representation includes both the attributes of each posts based on keywords and hashtag and the attributes of each audience, such as their ike, comment, share, and other attributes which reflect audience’s interaction with certain topics and certain posts. These inputs are then recorded by the system as data and sample and ranked according to relevance, therefore calculate the values to estimate by using the sample and data generate the output which has a numerical score that is a measure of how much the system believes that a particular audience will enjoy a particular post.

This isn’t the first and only time that Instagram has offered recommended content. One of the previous Instagram recommended content is subjective. You’d have to head to the Explore section to see the recommended posts and videos. Since they won’t be pushed to your home feed, the chances to see the recommendation posts depend on audience’s subjective choices totally. Like the change of switch from chronological feed to the algorithmic one and the introduction of ads, these are all based on machine learning and aim at broadening more content to uses and avoid missing interesting or crucial posts customized for each user.

 

Reference:

Alpaydin, E. (2016). Machine learning: the new AI. MIT Press.

Boden, M. A. (2016). AI: Its nature and future. Oxford University Press.

Kaplan, J. (2016). Artificial Intelligence: What everyone needs to know. Oxford University Press.

Josh, C. (2018). How Instagram’s algorithm works. Retrieved from https://techcrunch.com/2018/06/01/how-instagram-feed-works/

Kaplan, J. (2016). Artificial Intelligence: What everyone needs to know. Oxford University Press.

Sarah, P. (2018). Instagram will now add ‘Recommended’ posts to your feed. Retrieved from https://techcrunch.com/2017/12/27/instagram-will-now-add-recommended-posts-to-your-feed/