Categories
General

Steal These Ideas!

Talk is cheap, as the saying goes. That’s a good thing, since I am always overflowing with ideas that I have neither the time (I love my day job!) nor the money to advance. What I do have is a blog that I hope inspires readers to turn some of these ideas into reality.

My ideas are somewhat predictable, in that they all address user-centric information-seeking problems. Working for over a decade in this space has focused my intellectual curiosity somewhat — and of course I work on a number of these problems at LinkedIn. But there are many information-seeking problems that are outside of my present or foreseeable scope.

Here are two ideas that I’m hoping someone will execute on so I don’t have to:

1. Shopping: Help Me Figure Out What I Want

We’ve come a long way to improve the shopping experience, at least for utilitarian shoppers like yours truly. If I know exactly what I want, I usually find it by using Google as a gateway to Amazon, taking a bit more time if I’m feeling price-sensitive. I’d happily install a browser extension that could automatically detect product search queries and take them to my preferred shopping sites, bypassing the search results page, but that’s a minor detail of convenience (though probably not such a minor detail for the search engine companies). In any case, known-item search for online shopping is hardly inspiring as an open problem.

Exploratory search is another story entirely. For all the work that’s been done on faceted search, it is used almost exclusively to help people narrow search results. Progressive narrowing is great if you have a pre-established information need, but it is not the best interface if you’re hoping to evolve your information need through exploration. Instead of just “help me find what I’m looking for”, I’d also like to see more “help me figure out what I want”. I’d like to see an innovator applying faceted search to broaden queries, not just to narrow them, as well as going beyond collaborative filtering and “related items” to create a compelling browsing experience based on semantic and social navigation.

2. Organizing the World’s Information: Beyond Wikipedia and Navigational Queries

If shopping online often reduces to using Google to find product pages on Amazon, then informational queries similarly reduce to using Google to find Wikipedia entries. Nothing against Wikipedia — I think it is one of the most extraordinary achievements of our generation — but I think of the web as a library and Wikipedia as its encyclopedia section. Google’s mission statement notwithstanding, web search engines do a poor job of organizing the rest of the world’s information, instead choosing to optimize for known-item search.

There are countless opportunities for improvement here. Imagine if there were an interface for books, scholarly articles, patents, music, or videos that supported browsing and exploration of their content and meta-data. We’ve seen the beginnings of such an approach for individual libraries (e.g., the Triangle Research Libraries Network), but there is so much more to do in this space. Perhaps it’s a space that is hard to monetize, but even then I’d expect philanthropists to take an interest in making the world’s knowledge and creative artifacts more accessible.

If you are pursuing either of these areas, I’d love to hear about it. I’m sure readers here would too. I’m also curious to learn more about innovation in the travel and personals spaces, as those are both areas that could benefit from supporting exploratory search. And if you have work in progress, please present it at the HCIR workshop!

By Daniel Tunkelang

High-Class Consultant.

38 replies on “Steal These Ideas!”

Thanks for this! Great stuff as usual. I have a couple of brief reflections.

At my nonprofit, our mission is largely to help expose people to other opinions that they have not seen before. In this sense, I have thought a lot about my design task as an aspect of promoting “xenophilia” as it is described by Ethan Zuckerman. [1] I think his explanation of homophily goes a long way in explaining what you are ultimately trying to get at with your points about the limitations of the search interaction. I hope you will find the link a useful complement to your thinking. Contextualizing this work as exploratory search is giving me a brain full of avenues for design. I have worked previously on faceted search and absolutely have long seen it’s under-utilization in areas of exploration.

My experience in the for-profit world also makes me nod along to your assessment about the shopping experience. How many times will corporations pay usability researchers to remind them that yes people want to *actually shop* before they buy something? Many times optimizing the shopping experience just becomes a design for the cattle chute. The flat shopping experience leaves me without a satisfying experience — just a bunch of chutes to go down. Ultimately I think this lack of exploratory options often represents a distrust of the user and a lack of confidence in the product you are selling. Presently I think Apple is one of the worst purveyors of this greedy style of design, except the lack of exploratory affordances is aggrandized as chutes of “Genius.”

[1] http://www.ethanzuckerman.com/blog/2008/04/25/homophily-serendipity-xenophilia/

Like

Hi, Daniel. Can you help me understand the shopping idea a bit more?

Seems to me what you want is more of a search-as-a-dialogue interface? An agent that you can say something like, “I need a good, reliable vacuum cleaner”, and it responds with options that are highly reviewed, all good possibilities, and designed to better help you understand the options and elicit additional preferences? And then you go back-and-forth with this agent, refining your preferences, until you get down to a couple ideal options?

Is that right? Or, if not, can you elaborate more?

Like

Chris, thank you for the kind words and for your reflections. I’m a huge fan of xenophilia — in fact, I was hoping that various ideas by Miles Efron and Craig Newmark would encourage it. I regularly read a few blogs to get perspectives that challenge my worldview — whether about politics or software patents. But there’s no question that homophily is the path of least cognitive resistance.

In any case, I’m delighted that I’ve given you a framework that might help you in your mission. I’ll be curious to hear about your progress.

Like

By Post Author

Greg, here’s a simple example. My headphones recently broke, so I needed to buy replacements. I went to Amazon and searched for [in ear headphones]. Refined to the category Audio Headphones and then to Earbud Headphones and then to New. Lots of choices, and the facets don’t mean much to me. So I end up going to CNET to learn about in-ear headphones. I come back to Amazon again, this time looking at the Monster Turbine High-Performance In-Ear Speakers (which I ultimately purchased). The closest Amazon gave me to exploring the space from that example was “What Do Customers Ultimately Buy After Viewing This Item?” and an opaque “Explore similar items“. I ended up reading a lot of reviews, asking a few people I trusted, and then going with my gut.

I like the headphones, so perhaps I should be satisfied with this shopping experience. But what I would have liked is a process that educated me about the trade-offs among alternatives and helped me pick the right one. And yes, that probably involves going back-and-forth with an agent, refining my preferences, and ultimately understanding why a small set of candidates represents my ideal options.

Are there shopping sites that do this today?

Like

By Post Author

No, not that I know of, but I’ve been convinced for a decade that search-as-a-dialogue is badly underdeveloped.

I think it is true in almost all cases that people do not fully understand the range of options available, so need more information to be able to fully specify their preferences. That is the essence of the example you gave. And it is an extremely common pattern, as evidenced by the heavy use of multiple searches on a task and behavioral economic research that shows that people often have incomplete information and preferences when they first start trying to make a decision.

I worked on search-as-a-dialogue in the past (though it was more than a decade ago) and would love to see more work on it. It is remarkable how little progress we have made there and how many search engines still assume search is a one-shot deal (even ignoring the most recent previous searches when processing a new search) where people initially have complete information and are able to fully specify their preferences. This is a big, important, and relatively unexplored area.

Like

Daniel,

Nice posting. I totally understand the ideas vs. time dilemma.:-(

> If you are pursuing either of these areas, I’d love to hear about it. I’m sure readers here would too.

I know you are aware of this project, but in case anyone else that is interested, I am at the “baby steps” stage of creating a semantics-based graphically-browsed not-for-profit educational website and underlying knowledge base. (It could be extended to other domains as well.)

Preliminary details can be found here:

http://sites.google.com/site/rickcreamer/Home/miscellaneous-items/national-educational-software-technology-project

A very short proposal I submitted to the MacArthur Foundation is included in the page above.

Regarding shopping – yes, I’ve always wanted to have LEDs on the grocery store shelf light up when I speak into my “smart” shopping cart the name of the item for which I’m searching. If the item is not nearby, I’d like the LCD display on the cart highlight the location of the product on a store map display.

Like

You have very well put the conundrum about the online shopping behavior based on known-item vs exploratory search. Sometimes you just can’t express precisely what you are looking for and wish there was a natural language interface. That is precisely what we are focusing on at iApps.in for the App Store shopping and observing the search patterns I can fully concur the need for exploratory search.

The “known-item search” is very well suited for syntactic search technology which is a commodity now. PageRank is of little, if any, use in e-commerce. For truly “exploratory search”, we need a robust semantic technology along with temporo-social modeling. Adaptive facet presentation (brodening the query space) is also an excellent idea that can benefit from the semantic and social technologies.

Online search and discovery will continue to improve as more innovative approaches are explored.

Like

The “help me figure out what I want” is an excellent question! I have seen it often that when we have too many options, we get lost in what of the options are important and which are not. (“Do I want in-ear headphones, or over-the-ear?” “Is weight important or not?” etc)

In a selfish manner, may I suggest that the paper by our PhD student Beibei Li that will be presented at WWW2011: “Towards a Theory Model for Product Search” http://pages.stern.nyu.edu/~bli/www2011.travelsearch.pdf

The basic idea is to try to understand the preferences of the consumers and try to find the “best deal” for them.

We start by assuming that we have a (large number of) facets/attributes available, but not all of them are important for the user. Some people care about some features, some people care about others.

For example, if you search for a hotel and you go for a honeymoon, you do not care about the proximity to conference centers or about a business center in the hotel but you care about the beach. While if you go for a business trip, you may care about proximity to transportation and business facilities and beach becomes secondary.

By using techniques from demand-estimation, we estimate how much a user would be willing to pay for each of the features, (e.g., “a PhD student traveling for conference is willing to pay $2/day more to get ‘free’ Internet”, “a faculty member is willing to pay $5/day for the same”). Using this information, we compute a personalized “break-even price” for each product and each consumer. By comparing the actual price with the hypothetical break-even price, we can locate the best deals for a user.

Given the framework, we can guide the user in the selection process. For example, we know which attributes are important to different customer groups. So, when the user starts searching for a hotel but have no idea what to look for, we can ask the user to give us information about their user type (e.g., “I am a 30 yr old male, traveling for pleasure”) and show to the user what other, similar users cared about. (“People like you, place emphasis on proximity to the beach and walkability of the area”). If the user does not want to explicitly reveal their type, or when the users cannot even tell their own type, we can ask a few questions about their preferences (“what is more important? Beach or free internet?”) and infer their type.

This gives some structure in the search space, by identifying which facets are important for different types of users. Extra bonus: The technique is privacy-friendly and does not require data on individual user behavior to work. Having aggregate demand statistics is enough.

Like

“I worked on search-as-a-dialogue in the past (though it was more than a decade ago) and would love to see more work on it. It is remarkable how little progress we have made there and how many search engines still assume search is a one-shot deal (even ignoring the most recent previous searches when processing a new search) where people initially have complete information and are able to fully specify their preferences. This is a big, important, and relatively unexplored area.”

I completely agree, Greg. But didn’t you say for a long time that users are too lazy to have dialogs with their search engines? That people actually prefer the “do it all for me” approach?

I remember never accepting that viewpoint, because while it might be true for known-item search, I think the motivations and the desires and the needs that a user brings to an exploratory search session are stronger/different, enough to overcome that laziness. Just like in the shopping example above, sometimes people do just like to wander the aisles, even though it takes more effort to do so.

But having worked on it at an Amazon scale, do you still see that laziness as a large factor/problem/issue? And/or what do you see as the relationship between laziness and smarter interfaces? I.e. does a better search-as-dialog counteract laziness?

Like

@Panos, that paper looks interesting, thanks for pointing to it. I’ll definitely give it a read.

@Jeremy, I still think you want to get out of people’s way when they know what they want (esp. with re-finding and navigation), just let them get to what they want as fast as possible in that case. We’re talking about another common case, where people don’t know what they want, don’t know what is available, or both. In that case, exploration tools like browse categories, faceted search, and recommendations can help, as could search-as-a-dialogue.

What I like about search-as-a-dialogue is that it integrates easily with existing search user interfaces. I think the main difference is that you start by assuming that people will issue multiple queries per task and expect them to refine instead of treating every search in the session as independent. But I think you still probably want to present some results to the very first search, both because you might get the answer right away and because it helps people get more information about the range of options and their preferences.

Like

@Greg: Yes, totally agree with you on re-finding and navigation. If there is a known item search, why would you ever need (or even want) knobs to twiddle and buttons to push?

But all the knob twiddling algorithms over the years, all the grokkers and all the scatter-gathers and all the crazy hyperbolic results browsers (e.g. http://www.cs.kent.edu/~jmaletic/cs63903/papers/Lamping96.pdf).. those are all designed not toward known item or refinding, but toward exploratory, “dialogued” search in some form or other.

Speaking of starting by issuing multiple queries and then refining, let me self-promote for a moment and point you to a short paper that Gene Golovchinsky and I did last year, that’ll let you do exactly what you propose… get results right away and then build. Slides also available:

http://palblog.fxpal.com/?p=4453

Like

BTW, It’s interesting to realize that you had worked on search-as-a-dialogue before, Greg. I guess I never fully knew that. I went back and did some (now) known item searches on your blog, and realize that you posted a lot on that from 2004-2006. I started reading you in late 2006. Hence the gap in my knowledge.

Sometimes I feel, though, that the 00’s were the lost decade for exploratory search. A lot of what was done in the 90s falls into the category of what we’re now calling exploratory. And you yourself worked on it a decade ago. But the web came along in full force, and 95% of the research community’s interest shifted to known item and navigational search, as well as to their younger cousin, question answering. Exploratory search and recall-oriented search went MIA.

I’m encouraged that the pendulum is starting to swing in the opposite direction again. Now, only 90% of the community is interested in nav and QA 🙂 But wow, web search really took a toll on what I consider the more fascinating area of IR research.

I don’t have a point, I’m just lamenting out loud.

Like

Thanks, Jeremy. Those slides were interesting. I think I saw a version of that work before, either a paper, demo, or slides, but good to take a peek again.

It’s old, but the work a decade ago I did on search-as-a-dialogue was the Automated Travel Assistant, a search-as-a-dialogue for travel search. It tried to get a little bit of information about what flights you wanted, then showed you diverse pick of a few options, then encouraged you to add further constraints, and iterate on that, adding more and more constraints each time. It had a demo that used real, live flight reservation data, fun stuff. The paper has a couple hundred citations, so looks like some found it useful. Anyway, here’s the paper:

http://scholar.google.com/scholar?q=automated+travel+assistant

As for web search, the last thing I was working on at Microsoft before I left was a demo combining features of search-as-a-dialogue and personalized search for web search. Didn’t finish it before I left, sadly, but I think there’s a lot of potential there. Once you remove navigational and re-finding queries, the majority of sessions require more than one query to complete. Iteration happens.

Like

Folks, thanks for the links — looks like I might be the one stealing ideas after all! 🙂

But this discussion leaves me wondering more than ever: why is the market so resistant to exploratory search? Do users not want it? Does no one know how to monetize it? Is there simply a prejudice / status-quo bias that keeps people from trying?

Like

By Post Author

I think there is quite a hurdle for exploratory search. You have to design a complicated interface that remains highly usable, which is always a challenge. It cannot be slower than normal search for common one-shot and navigational queries (it has to stay out of your way when you want it too). It has to be immediately and obviously useful to users. And, even if you nail all of that, what you have in the end often requires some amount of training of users to understand, which is never a good thing and interferes with usability and perceived usefulness.

Not that it cannot be done, just that it is a hard task and no one has nailed it yet.

I wonder if a lot of things in the research stage go too far on exploratory search, though, in their effort to be novel. I think what is more likely to succeed is normal search with exploratory features added when it makes sense, much like the more successful attempts at personalized search have been normal search that included personalized features (like result reordering and targeted query refinements).

Like

I concur on your last point. It’s hard to make a splash with small interface changes, but radical changes often alienate users regardless of their utility. It’s a catch 22. The only winners are established players who evolve incrementally, and they often seem slow and resistant to change themselves. Perhaps rationally so.

Like

By Post Author

@ daniel et al re: “why is the market so resistant to exploratory search? Do users not want it? Does no one know how to monetize it? Is there simply a prejudice / status-quo bias that keeps people from trying?”

I don’t know if spoilt is the right word but Google has set the expectations of a search engine for the masses. Anything more difficult or requiring more steps than entering keywords (with autosuggestion) won’t be accepted (by the majority – Jeremy!). Google has also decimated the notion of charging for online services but that is another discussion.

Daniel mentioned patents as one of the areas under organizing the world’s information. The number of patents is increasing annually and more so with both China and India ramping up patent output. A key use for patent search is to find a set of similar patents and then focus in on the key ones. This is a known-problem that is talked about widely in the patent literature and industry. As you know, we built a demonstration patent search service to find similar patents and it received quite a few compliments from within the industry as it did what it said on the tin. But, there was little take-up from the organizations asking for such a capability. After some head-scratching I came to the conclusion that large organizations use public channels to vent their frustration with their existing vendors (which is a well known tactic) knowing that it is difficult to chuck them out but a way to force the vendors to offer solutions. Secondly, I noticed that even though standard text-search delivers poorer patent search results, people continued to use these services because it was what they were used to.

Like

FWIW, taking place at the Google booth in WWW 2010 this Friday:

http://research.google.com/www2011/

Quote:
“Search: Beyond static content
Kavi Goel
Mark Meiss
When someone searches on Google, we strive to give them the best results, as quickly as possible. However, there often isn’t one simple, correct answer, or the query isn’t quite clear yet. Instead, a user must issue a sequence of queries and click on a variety of results to explore the space. We demonstrate some of our efforts to create a more interactive and fluid experience to help people find information quickly and intuitively, whether searching web pages, finding driving directions, or even exploring the human body.”

Like

I’m still scratching my head in wonder as to why they don’t list music information retrieval as one of the areas to which they apply this “Search: Beyond Static Content” (aka exploratory search) approach.

Think about it: “However, there often isn’t one simple, correct answer, or the query isn’t quite clear yet. Instead, a user must issue a sequence of queries and click on a variety of results to explore the space. ”

That’s exactly what people are after when they’re exploring music. Sure, every once in a while you hear a song in the pub and want to do a known item “what was that song” or “who is that band” search. But more often than not I have an ill-formed query when searching through music.

I’ve been telling Google this since the fall of 2002. And yet it sees more value in exploratory search of the human body? C’mon.. the market for music is much bigger!

Like

Not going to post any links here, but I do think that a lot of users are interested in “exploratory search of the human body”. 🙂 Not sure how music competes with porn in terms of market size.

But more seriously, I agree with you — music is a ripe area for exploratory search.

Like

By Post Author

@jeremy
What do you mean by “But more often than not I have an ill-formed query when searching through music.”? What would you want from an exploratory music search site?

It seems to me that most people learn about new music from their circle of friends (or their kids). I hear a lot about people wanting better music search but not sure what the core need is. Thanks!

Like

“But more often than not I have an ill-formed query when searching through music.”? What would you want from an exploratory music search site?

I would want one in which I could “find similar” music by any number (and combination) of reasons that I chose. Some reasons include the things that Daniel mentions…finding global-top bands, global one-hit wonders, etc. Others are more data-driven, such as by friend-popularity (social music), by content-similarity (e.g. similar rhythms and/or chord progressions, as in 12-bar blues), or by other information…in addition to genre and language, maybe also country, decade, timbre.

Perhaps even by something like use of microtonality.

And the point is that I wouldn’t know about microtonality until after I started exploring. That’s (part of) what I mean about having an ill-formed query.

I don’t know what the right interaction model is, either, as Greg says. But music has such a richness to it that it’s hard for me to imagine that the only way people really want to explore music is in this sort of Google web navigation manner, “type in the name of the band and hear three songs”.

Like

Or maybe I even want to be able to explore music by friend *dis*popularity. I.e. help me find things that sound interesting to me.. that my friends are *not* listening to. So that I can be the cool one that introduces new bands to my friends.

You still need the social music graph in that case, but the way you are using it is not to boost in your recommendation what you’re friends are listening to, but to boost what they are not listening to, subject (of course) to other features and constraints.

So many fantastic ways to be exploratory with music.

Like

@Daniel: Exploratory search of the human body. Yes. Very funny. Very, very funny 😉

But wouldn’t your system realistically be exploratory search of multiple human bodies, not just the single Google body? 😉

Like

A really simple idea would be to take the metadata at http://www.allmusic.com/ and use this as the basis for an exploratory music search engine. Indeed, I’d think that a mashup of All Media Guide with the music and video content on the web (e.g., on Vevo or for Vevo content on YouTube) to support exploratory search would be a hit.

The best example of exploratory search for media I’ve seen is Netflix. For all of the hype around the Netflix Prize and optimizing for RMSE on its scalar ranking, Netflix doesn’t get enough credit for supporting exploration. There’s room for improvement, but I’d say its interface is about as good as it gets for a mainstream site.

Like

By Post Author

@Jeremy, I like your music search suggestions.

Just as image repositories (such as Google Images) could be analyzed and indexed both for embedded subjects, location, and multiple image processing statistical and morphological metrics which would enable you to then search for “mountain scene with blue sky and a lake,” music could be similarly indexed – especially after you assert you like one particular recording (or a specific part thereof such as the guitar solo on, say, track X on album Y recorded by artist Z). Or, as you mentioned, assert examples of what you do not like.

Then multiple websites could be dynamically searched to: 1) perhaps find a downloadable MP3 file of the recording you specified which could be DSP-analyzed (by several servers in parallel) for various attributes/metrics such as genre, tempo, time signature(s), key, drum part category (perhaps even a leaf node in a simple percussion taxonomy like “double base drums, up-tempo, rock genre, circa 1980”), instruments used, and 2) from other fan, magazine, or artist sites, determine who the individual (e.g., solo guitarist) musician was, which amplifier and signal chain he/she was using for the specified recording to characterize/identify, say, a specific guitar tone/tempo/style combination that you like (e.g., Pete Townshend’s Hiwatt amplifier with a Gibson SG).

Then, given the rich, multi-dimensional set of metadata describing the one musical piece you like, search for and return similar patterns, i.e., attribute combinations. For instance, other recordings in which the same artist performed at the same tempo and general key category (major/minor), perhaps in the same musician but in different groups, or other artists using similar signal chains, tempos, keys, time signatures, genres, etc.

Perhaps even as simple a UI as one or two columns of attribute/value check boxes would enable a significant improvement on the search results presented to users. Or perhaps a fill-in-the-blanks phrase oriented UI (with combo boxes in the blanks) might work as well.

Like

Richard,

You are spot on with your observations. So much so, in fact, that there is a company doing this: http://the.echonest.com/platform/

I’m just wondering why search companies whose explicitly stated goal it was to organize ALL the world’s information haven’t been doing it, and have instead been working on chat programs and spreadsheets.

Like

TechCrunch post today:

Google's Music Search Engine Quietly Vanishes From The Web

Here’s the key bit:

“In a blog post, Google said that, going forward, a simple Google web search would enable users to “search and more easily discover millions of songs”. Queries for songs, artists or albums would return search results including links to an audio preview of those songs provided by its music search partners, at least in the United States – for starters.”

That’s the part I object to.. that they thought that the best way to “explore” music would be for you to type in an (already known) band name to hear a few songs.

Let’s hope that in the future, the people working on music and the people working on Google Body (or whatever it’s called) start talking to each other. Because I’d really like to see a panel announcement that looks like this:

“Search: Beyond static content
Kavi Goel
Mark Meiss
When someone searches on Google, we strive to give them the best results, as quickly as possible. However, there often isn’t one simple, correct answer, or the query isn’t quite clear yet. Instead, a user must issue a sequence of queries and click on a variety of results to explore the space. We demonstrate some of our efforts to create a more interactive and fluid experience to help people find information quickly and intuitively, whether searching web pages, finding driving directions, exploring the human body, or even listening to, interacting with, and enjoying music.”

Like

@Jeremy, they might also offer users the option to upload their own mp3 file they like along with optional time offset and timespan describing the part of song one likes. They could then analyze the mp3 before showing an attribute preference-oriented UI.

Thanks for the link – I was unaware of it.

Like

I wonder if an mp3 analysis engine could pave the way for a paid copyright infringement search service?

Like

@Richard: I actually pitched that idea to Google (music content-based copyright infringement service) in mid-2001. But even then, the idea wasn’t new. Yamaha Research Labs had done it in the mid 90s, and Fraunhofer, Nokia, and Shazam had all done it by the late 90s/2000.

Like

Though I’m late to the shopping conversation, I’ve been using http://thefind.com as my startpage for product search quite extensively for years. I find it works much better than Google and offers me the ability to search locally also. No, it doesn’t offer the full exploratory aspect of known/unknown unknowns in terms of discovery, but it’s dead focused on a better ecommerce search experience.

Like

Comments are closed.