Relevance is one of the most sought after qualities for any kind of content presented to a user. Clicks are a user’s declaration of relevance within our conventional browser UX for presenting content, media, news, and information. Relevance becomes a critical tool for navigating information as the amount of information available on the Internet grows every second. Measured in terms of physical size, we are approaching ‘Nebula‘ scale, measured in lightyears for the amount of information available on the internet today.
Presented with this interstellar scale of information, the challenge in navigating the mass of information is no less daunting than figuring out information-entropy (Shannon entropy) of online content filtered/ranked by relevance. Put another way, all information is high entropy (high uncertainty) unless sorted, presented by relevance to each user.
Ideally, all content should be customized such that the presented content is relevant to that particular user and more likely to engage the user without requiring inefficient actions (searching, clicking, scrolling) on the part of the user or requiring a constant declaration of relevance on every tab and every site.
In this post, I attempt to understand how relevance works for content recommendations on some of the big sites and examine applying this model of determining relevance to commerce online.
In the context of relevancy, the 2D wall of content at Pinterest is very conducive to more browsing but not necessarily deeper engagement as content is not sorted (or sortable) by relevance. While categories help, they are not personalized and presentation of content does not sense any other evanescent declaration of user interest on their site or other sites. Thus, it is clear that:
Relevance != Social Proof
Quora’s UX surfaces high quality but not necessarily high relevance by relying on democratic votes (1 vote per user regardless of their Quora-ranking/clout/points). As their product evolves, I hope we will see more relevance – perhaps by picking topics as well people of interest. This is a form of declared relevance relying on declaration made by the user, not necessarily determined by a user’s ongoing interaction with content on the site. The takeaway:
Relevance != Quality of Content
Twitter’s #Discover tab is the most advanced relevance-filter yet with its mix of tweets by people you follow as well as people who share your interests as determined by cookie tracking on sites you visit + interests of people you follow. I think this is the closest determination (ongoing) of relevance in an online product.
Relevance = A mix of [Social Proof + Quality + Personalized-Interest matched Content + ...]
Relevance and Commerce: curating for serendipity
While Google, Twitter, and to some extent Quora address relevance for the world’s entire corpus of information, e-commerce companies can and should apply similar techniques to determine the right products for the right buyers. While narrowing down choices presented to a potential buyer or browser, the content selection should permit some amount of delightful discovery. This need for counterbalancing narrow, algorithmic selection and presentation speaks to the emotional part of a user’s browsing experience.
Paraphrasing Tufte, all design is choice – and a failure to engineer correct information flow underlying presentation [balancing relevance vs. discovery] will surely result in clutter and confusion. The balancing act for narrow relevance in commerce is serendipity – let the user discover adjacent information/products albeit in a delightful way. And thats where I think we’re headed in commerce – augmenting relevance with sufficient serendipity to deliver the right user experience. While algorithms do well with determination and tracking social proof, quality, personalization using user-interests, serendipity requires blending in a measure of user and domain-expert curation. And, therefore
Relevance + Serendipity = Algorithms & Machine Learning + User curation + Expert Curation
Within online commerce, there have been three broad waves of innovation:
Wave 1: Digitization of product information, browsing, and fulfilment
Amazon is clearly the best example of this class of innovation with its broadly horizontal digitization of product information followed by a simple layer of product-interest matching and recommendations. If you saw product X, you may also be interested in product Y where Y may be the most viewed/purchased item along with X. There are no social or other user-level connections that Amazon seems to use other than the user’s history on its site. Currently it presents six items in at least six categories if I visit the homepage (logged in). It hasn’t even asked me a Hunch.com style wizard to narrow down ‘recommendations’ or increase relevance to me.
Wave 2: Propelling discovery by economic compulsion (ok, by surfing the curves of indifference)
Groupon and other daily-deal sites induce discovery by providing an economic compulsion for users. These intermediaries harvest a user at their point of indifference in the face of compelling economic value for the presented product. With enough data points, Groupon et al will have their own version of consumer demand curves. I expect they will move beyond offering a single daily-deal towards a smorgasbord of carefully chosen goods that are seen as acceptable substitutes given a certain budget constraint. They certainly have vast consumer purchase data to do so within local commerce – a valuable dataset vs. Amazon which is purely online. Groupon Goods, I hope is the first move in this direction. If I was at Groupon, I would hire Economists as well as statisticians and data-scientists to figure out indifference curves and match it to a variety of local commerce.
Wave 3: Guided discovery – engineering serendipity
Faced with an unprecedented data storm, consumers need/want fewer choices but the right choices. Balancing this need for narrow personalization is serendipity. Serendipity is pleasant and welcome because it helps users make a useful discovery even though they were not explicitly not looking for it. This middle ground is guided discovery. Lets examine how information/content sites deal with engineering for serendipity.
Serendipity doesn’t happen on a Google search results page as a user (and PageRank) explicitly rules it out in favor of surfacing the information users are searching for from a sea of information.
Amazon recommendations hint at some serendipity but is strictly dependent on user’s previous purchases, items browsed, and most popular items related to a user’s purchases. The form of guided discovery on Amazon doesn’t leverage a user’s behavior or interest-profiles that exist elsewhere.
Pinterest at this point (May 2012) is random serendipity which is not very time-efficient pursuit for the user though it yields great time-on-site and other vanity metrics for Pinterest. In effect, there is no incentive for Pinterest (yet) to boost relevance vs. pageviews. Guided discovery only takes place along canned categories or along content classified by the users in various boards.
Quora in some ways is guided-serendipity by walking the user along the axis of quality content. The only axis of serendipitous discovery is following people and their Questions/Answers/Boards/Posts. If one could do Quora score/votes based recommendations for products/brands and integrate some level of sentiment analysis of Quora answers, it could become a compelling front-end for products with better insights vs. Consumer Reports.
Twitter brings together some nice elements of guided discovery by mapping interests, people, and recent events/topics of interest to your chosen geography at city/national levels of the twitterverse. I think Twitter is ideally placed to guide users towards all kinds of media in addition to news and discussions. It can help me find media, content, news, and information based on:
- Interests I follow (on/off Twitter)
- People I follow
- Interests of People I follow
- Activities of People I follow (people followed by them, their retweets, favorites, …)
As far as I know, there are no equivalent efforts in online commerce applying data mining based recommendations to guide the users towards the right mix of guided and serendipitous discovery. Merely suggesting some ‘recommendations’ in a side-bar don’t suffice. What we need is a relevance based content-display and navigation system.
This I believe is the next big wave of commerce – data curated commerce to help the users browse less, find more.