Recommendation Systems

The Bikes not Binkies reference is something I thought of as I was monitoring the great Twitter conversation during yesterday's very successful AMA webcast on Personalization featuring our head of marketing, Carlos Carvajal. 

First generation personalization systems relied very heavily on a user's past behavior or tried to box them into a particular category.  For me, this box, is always whatever I was last buying my kids.  So while my 3 and 4 year old are ready for bikes, some sites, which will go un-named, are still trying to sell me binkies, in otherwords baby stuff.

Personalization is really a strategy, not a technology.  It's a component of a broader customer experience optimization strategy.  A personalized site is a site that works for the person who visits, i.e. different for me in my context of "mom of toddlers" than for my context "daughter needing to buy mother's day present."

There were so many nuggets that people pulled from the webcast and posted on Twitter, frankly some of these were so cogent, you might see them pop up in other Baynote material.  Sometimes being limited to 140 characters is a really good thing.  Ok, here is what some people tweeted:

"Personalized landing pages are a simple & effective way to get started with online personalization based on context & intent" 

In case you don’t know Bluefly, it is one of the hottest places to shop online for designer fashion. Bluefly was featured last week in the National Retail Federation’s magazine STORES. The article was titled “Power of the People: Collective intelligence translates into increased sales for Bluefly.com.”

The article described having Baynote Recommendations on your site as like having a “retail floor teeming with hundreds of associates.” You know, the helpful and knowledgeable kind, the ones who can show you exactly what you are looking for.

Marty Keane, the company’s senior vice president of e-commerce, talks about how they achieved 300% lift in sales by harnessing the wisdom of all Bluefly visitors. One cool angle on the Bluefly story is that, as a customer with a rapidly changing inventory, it was imperative that their recommendation system have no lag. Because Baynote recommendations are based on the preferences of all site visitors, not just purchasers, Bluefly’s “most wanted” recommendations on new arrival product pages keep pace with the changing inventory.

The final installment in the 4-part series, “My Search Sucks,” discussing why search, well, sucks.

Over the past few weeks, we’ve explored how there are three key principles that explain why site search just doesn’t perform like we expect it to and what we can consider to help mitigate this.  So far, we’ve learned that:

1. The critical information we need to make search great isn’t in the document – it’s in the users’ heads.
2. Asking users to explicitly provie us information that would improve search, while a seemingly good approach, is inherently flawed.

Today, we explore the third principle that shows that if we want to improve search, we need to focus on all of the things users are doing online.  What I mean here is that we need to look beyond search and at the entire site experience to truly understand what’s valuable, why it’s valuable, and in what context it’s found to be valuable.

Reason #3:  Search does not exist in a vacuum.

In order to improve search, we need to observe more than just search behavior.  Search and navigation have traditionally been seen as two separate paradigms: separate interfaces and separate systems driving them.  But in reality what’s happening?  A user is coming to your site and expressing an interest or intent through their actions.  They might have first expressed that intent through a Yahoo! or Google search that brought them to your site.  They might then express it in the pages they visit and engage with, the navigation they use, the links they click, and maybe the site searches they perform.  This expression of interest may span multiple searches and clicks.  And, finding documents that hold true value for that interest and intent may also take multiple steps.

Let’s think back to the “Insight/Incite” example once more.  Had we only looked at what search results users clicked on, the problem might never have been solved.  Why?  Because the valuable content was never in the results – it wasn’t there to be clicked on in the first place!  To learn what users really meant by “insight”, we had to watch their subsequent navigation, paying particular attention to the patterns of behavior that indicated engagement or that they had discovered content that was of value – even if it happened several steps after the initial search.  Observing search behavior alone is not enough!

What about users who don’t search at all?  What can we learn from them?  Users are actually giving us continual clues to their intent and interest with every link they click and every category they choose.  The documents that users engage with and the order in which they engage also tell us not only about relationships between documents, but intent.  If we take this valuable, implicit insight into account, then we really begin to see how this insight could be used to fix search.

What’s really remarkable is that once we take a step back and think of the entire online experience as a single unified expression of intent and value, we can do a lot more than fix search.  We can start to make recommendations and optimize the user experience with every interaction they make with your site; from the moment they arrive, every step they take through the site, as well as every search they perform. The true goal is to understand the user’s intent and then automatically surface documents that other like-minded peers have found valuable in that same context.  That’s the true wisdom of the crowd, and what Baynote’s Collective Intelligence Platform (CIP) is all about.

The central focus of Gartner’s Symposium/ITxpo this week in Orlando is all about implementing what they’ve recently dubbed as a “pattern-based strategy”.

According to Gartner, a pattern-based strategy “provides a framework to proactively seek, model and adapt to leading indicators, often-termed ‘weak’ signals that form patterns in the marketplace.”  For the past several years Baynote has been committed to helping companies identify these patterns with technology that lets them tap into the collective intelligence of customers visiting their websites. This is something that transactional based systems such as business intelligence (BI) and complex event processing (CEP) simply haven’t been able to deliver. Here’s why:

1) For years BI, CEP (more recently) and other related technologies have helped organizations become much more efficient by automating their interactions with customers. However, in the process of creating huge economies of scale, they forced companies to lose the “mom and pop” touch that consumers expect when they walk into a local hardware store or restaurant. In failing to create digital mom and pop experiences, online retailers and publishers have placed unnecessary emphasis on promoting popular products and content, thereby losing out on profits to be gained from merchandising their long tail products.

2) In addition, these so-called “predictive” applications have historically prioritized the wrong set of indicators, often identifying consumer trends weeks, if not months, after the fact. For example, e-commerce transactions lag other more relevant indicators, such as online comparison shopping, by months. Only by tapping into the power of the collective is it possible to see early signals, spot trends and develop strategies around them before your competitors catch on. This holds particularly true for long tail products. Our customer US-Appliance tapped into the implicit behaviors of its website visitors to merchandise colored washers/dryers months before Home Depot and Best Buy began promoting similar products in their stores.

In Gartner’s recent report, entitled “Introducing Pattern-Based Strategy”, they view “the collective” as being critical to developing a pattern-based strategy. We couldn’t agree more with their position:

The collective comprises individuals, groups, communities, mobs, markets and firms that shape the direction of society and business. The collective is not new but technology has made the collective more powerful — and enabled change to happen more rapidly. The explosion of social software has enabled groups and individuals to rapidly form and rally to a cause — often resulting in significant societal changes.

The result for business is a cacophony of rapidly evolving demands, expectations, inputs and transactions, as well as an opportunity to not only react, but to seek signals of change from the collective. Market trends, some subtle, others strong, are masked by noise, and many enterprises are failing to proactively detect the patterns they rely on to direct future strategy and support investment decisions. In addition to failing to detect these patterns, enterprises are not utilizing new resources to proactively seek signals of change nor do they understand their power to influence individuals and communities.

Val Sribar, group vice president of Research at Gartner, sites Amazon’s and Netflix’s use of recommendation engines as good examples of organizations leveraging collective intelligence to support their pattern-based strategies. Sribar agrees with Baynote that recommendation engines identify new patterns in behavior as customers browse and purchase. While Amazon and Netflix are highly popularized cases, we’ve helped hundreds of other well known brands tap into their collective customer networks to significantly increase revenue through cross-selling and upselling, and higher customer loyalty.

We’re excited to see Gartner take a leadership position on this important issue and look forward to working with them and our customers to bring best practices related to collective intelligence to the forefront of modern business strategy.

According to a new consumer privacy study by the Berkeley Center for Law & Technology at UC Berkeley, and the Annenberg School for Communication at the University of Pennsylvania, two thirds of Americans object to online tracking by advertisers. The study is apparently the first national telephone survey that explores Americans’ opinions about the controversial practice of behavioral targeting.  Here’s a statement from the press release about the report, which was issued on Wednesday:

The report, Americans Reject Tailored Advertising shows that 66 percent of adults said no to tailored ads. Not only that, when informed of specific behavioral targeting techniques that marketers employ to create the ads, even higher percentages — between 73 percent and 86 percent — oppose tailored advertising. Those techniques include tracking behavior on websites and in retail stores.

For a more detailed analysis of the findings, you should check out Stefanie Clifford’s coverage of the report in the New York Times.

We’ve been talking about the pitfalls of behavioral targeting for years, so it’s nice to finally see some national research that tells marketers what consumers actually think about this shady technique. In this age of identity theft and mounting concerns over privacy in general, a practice that proactively profiles a user — perhaps over the scope of many Web sites and over a period of several months — will sound alarms even among the least conservative of us.

Beyond privacy concerns, there are bigger issues with behavioral targeting related to accuracy and quality, that many marketers still don’t understand. Traditional behavioral targeting struggles precisely because it tries to discern what I want now based on my past behaviors. Consider the impact of focusing on historical interests instead of current intent: If I bought a gift for my niece on Amazon.com last week, I certainly don’t want to be bombarded by ads for similar products that probably aren’t relevant during my next visit.

Another way to think of this problem is to consider the idea of roles or what personalization systems might call “profiles”. Humans have far too many roles in life for a profile to possibly predict what a user wants on any given day. A woman shopping for baby clothes, a tie for her husband, and a gift for her sister may appear schizophrenic because she is acting in three different roles — mother, wife and sister. What do you show her next? Tossing strollers ads at her isn’t going to be effective now that she’s shopping for a new cocktail dress for herself.

This is the pitfall of profiles. In a given month, an individual will have thousands of roles. Knowing my past is not necessarily a better way to predict my future. In fact, this phenomenon has been known by psychologists and other scientists for years — humans are animals of context and situations, much less than of historical profiles or roles.

Enter Intent-driven Targeting

An alternative that solves the issues with both privacy and effectiveness is one centered on understanding the users’ intentions, instead of their clickpaths or profiles, and pairing that knowledge with specific content, product and advertising recommendations. This approach relies exclusively on the collective wisdom of like-minded peers who have demonstrated interests or engagement with similar content and contexts.

The concept of profiles is completely removed in this case. Instead, through understanding expressed or implied intent, content appropriate to the user’s current mindset can be delivered.

Most importantly, it kills two birds with one stone: Users get useful, accurate recommendations and ads, while still avoiding the whole privacy mess.

Richard MacManus over at ReadWriteWeb recently turned me on to an interesting YouTube blog post about the effectiveness of the popular video aggregator’s 5-star rating system.

The post, written by YouTube product manager Shiva Rajaraman, explains that the majority of YouTube users who rank videos give them a perfect 5-star ranking. He continues:

Seems like when it comes to ratings it’s pretty much all or nothing. Great videos prompt action; anything less prompts indifference. Thus, the ratings system is primarily being used as a seal of approval, not as an editorial indicator of what the community thinks about a video. Rating a video joins favoriting and sharing as a way to tell the world that this is something you love.

Rajarman goes on to solicit the community for feedback on how useful the current ranking system is and what can be done to improve upon it.

We’re really glad to see that YouTube is finally examining its rating system with an eye on delivering more value to its community and look forward to seeing how the system evolves from here. Ratings and user generated reviews, though often misleading, have become an expected part of the online experience and encourage deeper engagement. I don’t think anyone would take away points from YouTube on their ability to engage an incredibly large, diverse and influential community of users. However, YouTube’s review system- and others like it –  must also find ways to inform ratings based on valuable sentiment and implicit feedback gathered from the vast majority of their site visitors. Not the loud minority.

With a truly integrated approach to recommendations that blends both implicit and explicit feedback, companies can expect to improve engagement and overall user experience by directing site visitors to the best content based on their intent.  I talk a lot about this concept in my paper, entitled “7 Deadly Biases”.

In the end, explicit versus implicit user feedback shouldn’t be viewed as an either/or scenario. Please let us know your thoughts on the matter and share examples of sites that are doing it right.

Yesterday at Baynote we had the privilege of a visit by a delegation from China’s Consulate General office in San Francisco including the Consul General himself Gao Zhansheng.

We initially discussed Baynote’s platform and why its peer-driven recommendations are particularly important to business and society at large. This lead to a more philosophical conversation regarding crowd-wisdom and how out of touch both big business and big government can be. We discussed how so many of the specific problems which have lead up to today’s financial problems were matters of disconnect. You’ve heard me say it before but people want the connection they used to have with “mom & pop” businesses. Likewise, they probably also want the connection they get in “town hall” interaction with government leaders. The Chinese officials teased with an old saying in China: “connect the crowd and serve the people.” Except that we all want to do this online now.

Despite the difference in our economic systems and the fact that these weren’t technologists, the concept of “the crowd is free” really resonated with them. They understood the limits of experts. Someone on their team asked about the applicability of crowd-wisdom to government sites as a means to be more responsive. Except for NASA.gov, there are few U.S. government websites doing this….yet. It’s coming, it has to.

When I was telling them we had Baynote on roughly 180 sites, the Consul General joked “That’s too bad, I wish it were more, I have trouble finding what I need on the web, you need to get out onto more sites.” “We will, they are coming,” I responded.

“You know Chairman Mao understood this,” someone commented. “He said ‘Crowds have infinite power.’”

While we didn’t explicitly talk about the massive bailout announced this Sunday, it was clear to me by the tone of our conversations that the team from the consulate were looking at today’s climate as a global issue. They wanted to know what it would take for a business to make it in this climate, not just in China but here in the U.S, and what kind of governmental policy or innovative technology they can help to push.

I explained that it’s never easy, but that starting a business in this environment is incredibly tough right now. It’s winner take all — you would have to be #1 to make it here. The upside is that there are a lot of great lessons to learn from the wave of “nice to have” Web 2.0 technologies that can’t demonstrate ROI. Be innovative, address a real market problem, and be able to demonstrate ROI quickly with very little risk and you could thrive. Baynote is fortunate to be in the position we are in today, and much of our success comes from following these rules.

At parting, they presented us a nice set of gifts — a scale-down Terracotta soldier, a couple of DVDs for the Beijing Olympics… The Baynote team was truly impressed by the officials. They were smart, very knowledgeable, curious and extremely passionate about innovations and technologies — not normal traits of any diplomats. We wish them well!

This little doodle gives some insight into the inherent problems of using visitor profiles to target recommendations. We’ve all been at a site or two where the recommendations just don’t make much sense–what would happen in a real store using that same profile-based targeting system?