In the following pictures we show you different real-life examples of recommendations on some e-commerce and other sites such as Amazon, Flipkart, Times Of India, and eBay.
In the following picture, when we select the book Programming in Scala at Amazon, we see suggestions (or recommendations) for those books that were brought together with the selected book:
The following is another example of product recommendations based on the frequent patterns, that is, the products that are bought together by customers. Here Moto E mobile phone and Moto E back cover combo are recommended for the user to buy.
We can also determine that if two people like similar items, then if one person liked/bought something that other person didn't, we can recommend it. The following are some examples based on people to people correlation. See the header Customers Who Bought This Item Also Bought.
Today e-commerce sites are so popular that we tend to buy as much as we can online because it saves us time and energy. However, our decision to buy a product greatly depends on the reviews and ratings provided by other customers. Check out the Customer Reviews and Most Helpful Customer Reviews sections in the following screenshot:
Notice how the customers' aggregated rating (stars) and their distribution across five different ratings is shown. Also note that customer reviews are listed together, making it very easy for a user to make decisions.
In the preceding screenshot, we see no reviews for a given product; however, that space is complemented by recommendations based on the user's recent history. Basically, we fill in the space with some information that could help a user find relevant items.
We can use the search results of other users who have also searched for similar products and provide it to our current user. This will give the user a larger variety of options in products and also keep him/her interested in the site. The following screenshot illustrates this:
We can also add another dimension to the mix, that is we can recommend based on the current activity of all the users. The following picture demonstrates that on selecting a Dell laptop on eBay we see what laptops people are currently watching. This makes sense in those situations where we have a lot of people currently making transactions, and the lifetime of a particular item on the website is pretty short. This is true of consumer to consumer sites. On such sites, people auction an item, and it is sold in a very short time of maybe hours, days, or weeks. However, on many other e-commerce sites, the items would be available as long as they are available across the markets.
In the following screenshot, we see similar news items, which are placed right beside the current news item. This is an example of item to item correlation:
In the following table, we can summarize the three components of a recommender system. Notice how the recommendation technology is annotated with appropriate input data, below it. Also note that this is only a selective coverage of these websites, not an exhaustive coverage. There may be many more features that are not immediately obvious.
Site |
Recommendation interface |
Recommendation technology |
How user finds/reaches a recommendation |
---|---|---|---|
YouTube |
Similar item |
Item to item correlation Input: Previous videos watched |
Organic navigation |
Amazon |
Similar item Top N list Average rating Comments/reviews |
Item to item correlation People to people correlation Input: Purchase history, aggregated rating, and user reviews |
Keyword product search Organic navigation E-mail subscription |
Flipkart |
Similar item Top N list Average ratings Comments/reviews |
Item to item correlation People to people correlation Input: Purchase history, aggregated rating, user reviews |
Keyword product search Organic navigation |
eBay |
Comments/reviews Average ratings Similar items |
Aggregated rating People to people correlation Input: Likes / text / navigation history |
Organic navigation |
IMDb |
Similar items Ordered search results Average rating |
Item to item correlation People to people correlation Input: Aggregated rating |
Organic navigation Keyword search Categories |
TOI News |
Similar item Browsing |
Item to item correlation Attribute based Input: Editor's choice |
Organic navigation Keyword search Categories |
There are two more metrics we need to discuss. First is the effort put by a user in actually reaching to a recommendation on a scale of manual to automatic. Second is the persistence of recommendations on the scale of ephemeral to persistent.
Manual recommendation involves an active effort on the part of a user to locate the items he/she is actually interested in. For example, in the case of a news site, a user might need to browse or search for a news item before interesting news is reached.
Automatic recommendation is displayed to a user as soon as he/she visits the website, for example, when a user is browsing for USB storage, after buying a mobile phone the last time. If the website actually provides recommendations for the latest USB storage devices at this time, then it is a real time saver. The site will find a loyal customer in this user.
Ephemeral recommendations are generated for a single user session, and it doesn't really utilize a previous transaction history of a user. For example, if a news site simply displays the most popular news story of the week, then it is true for all the users on this site.
Persistent recommendations require that the likes and dislikes of the user are known in advance so as to make personalized recommendations to a user. The USB storage example that we saw earlier is a good example of persistent recommendation.
The optimization problem in a large e-commerce store is the maximization of profit and at the same time it is minimization of the in-store wait time for a product. The longer a product stays in a warehouse, the more its profitability decreases because other easy-selling products cannot be stored in the warehouse. It is easy to guess that frequent pattern suggestions work best when provided with a combined discount and customers prefer to see ratings and read reviews before buying products on-line. We can look at different sites from different angles and think about how to better serve the user with good recommendations.
Now that we have an understanding of which concepts of a recommender system we can leverage, in order to make the customer experience better, let's look at the architecture of a typical project.
18.116.65.130