Amazon recommendations worked for certain categories like books – people with similar tastes bought similar books and the set diagram approach (affinity analysis) to recommendations worked for them. However, Amazon expanded to include more items, various suppliers and it is time recommendations improve along the business.
Amazon has two sets of users – people who go there with a specific item in mind and want to find a better bargain for a particular item. Other group of users goes to Amazon with a specific category in mind, like Eau de Cologne or movies. They know that they want to buys something new but do not know which brand or item to buy. They are there to browse through and find an item – existing recommendations does not work for them.
Recently, I was after cologne or two. I was not sure what I want so I thought I would browse the store and find something interesting. In the home page, I see list of items I viewed recently. If I go to a category, then I see ‘Customers who bought items in your recent history also bought…’ I see lot of recommendations there but they are items I just browsed.
This is where the recommendation is failing – I like particular type of colognes but I do not see them in the list at all. It was not personalized to me.
Let us take, Terre D’hermes for example. It is woody spicy with citrus accord. I like it and I bought it last summer. Comparable colognes would be Allure Homme Sport, La Nuit de l’Homme and to some extend Acqua di Gio Pour Homme and L’eau D’Issey pour Homme. I provided these as an example because they are either Citrus accord or woody spicy / woody aquatic variants - chances of me buying these something would increase if I see any of these in my recommendation tab.
Instead, I saw recommendations that were totally off including ‘Paris Hilton for Women’ and ‘S by Shakira’.
Amazon has everything going for them: tons of products, tons of transactions and tons of people and purchase / behavior history of users. Is there a possibility to fine-tune the recommendations?
I think the key is to use the metadata. Though Amazon never exposes the tags, I was able to find it buried in the recommendations page. Recommendations page let’s you customize your recommendations and rate products similar to Netflix’s recommendations page. In there, I found Amazon’s tag associated with Terre D’hermes. I can add the tags to the product. They are crowd sourced and not perfect.
My biggest problem is that they make you do lot of work to set up your recommendation. Even for items from your previous purchase needs addition of tags so they could be used for recommendation.
Also, these tags should be part of the item’s inventory. It can include user tags but each items should have attributes as tags. If Amazon has it, it was not exposed. I am going to assume that it does and hypothesize how it can used to improve recommendations.
Based on my previous purchase history, t would be awesome to see a capped list of ‘earthly’ colognes because of my previous Terre D’hermes purchase. Generally people prefer heavy colognes for winter and lighter / fresh ones for summer. Based on season and purchase pattern, I could get different colognes or fine-tuned ‘earthly’ colognes selections based on different variables. It is one step better than just purchase history. Items with more peer reviews should be a variable too – more positively reviewed items should get higher weight.
Just for colognes, there is lot of variables and signals for a targeted recommendation. For clothes, kids shopping, etc., you need to find set of unique variables and signals for recommendations. Fashion shops / designers have unique style, people from different countries buy different styles of clothes, colors size, other clothing articles to complete your set, season, and even clothing cuts (European, regular, slim), etc. are signals for recommendation. If you use these against a person’s purchase history, the recommendation will get better.
With Kindle, Amazon is in a unique place to recommend targeted personalized options. They have huge data set and through Kindle more personal data that other online retailers might not even dream of. Shows and movie people watch, social graph – Facebook attributes like birthday, friends list, likes, check-ins, following, etc., can help personalize better. Kindle can leverage ‘X-Ray’ and recommend purchase options for certain show or events. For example, Oscar Red Carpet event can have purchase recommendation for original / similar accessories, clothes worn by the actors / actress.
A ‘Similar’ category should not only generate other clothing options but also lower / higher price points for a particular item.
Recommendation is science but also an art. It takes lot of data, interpretation of data, user behavior to get there. Personalized recommendations provides 10% boost to the sales per various case studies – even if revenue increases by 1-2%, it is a good chuck of change.
Amazon is a data driven company – they capture and analyze lot of data and experiment them to increase sales. I am curious whether they do anything similar to what I have outlined for recommendations.