Personalization in Search Engines Personification or Penalization A

Personalization in Search Engines: Personification or Penalization? ! : פרסונליזציה במנועי חיפוש !? האנשה או הענשה טלי שרון , אריאל פרנק A. Frank-T. Sharon

Blog sources as motivation/basis for presentation • “Problems of Personalization”, P. Lenssen, Google Blogoscoped, March 24, 2005, http: //blog. outer-court. com/archive/2005 -03 -24 -n 33. html • “Personalization is Hard. So What? ”, G. Linden, Geeking with Greg, March 24, 2005, http: //glinden. blogspot. com/2005/03/personalization-is-hard-so-what. html • “Search Engines: Near and Distant Future”, P. Lenssen, Google Blogoscoped, May 17, 2003, http: //blog. outer-court. com/archive/2003_05_17_index. html 2 A. Frank-T. Sharon

Contents • • 3 Google intelligence? People perspectives and privacy Over-simplified Personalization? ! Personalization and relevance rank Why is Personalization hard? Local/Locality changes (technical) People/Times change (fundamental) A. Frank-T. Sharon

Google intelligence? 4 A. Frank-T. Sharon

Personal talk with Google? (1) • Imagine the following conversation - not typed, but spoken, using speech analysis and speech output: – John: Where can I find a restaurant nearby? Google: Where do you live? John: In Manhattan. Google: What do you like to eat? John: Oh, something Chinese, but cheap. Google: Should I reserve a table at the China Wok Restaurant five minutes from your place, in one hour? John: That would be fine. Google: Done. 5 A. Frank-T. Sharon

Personal talk with Google? (2) • Can you see how much data the search engine now collects? Next time (if the user allows privacy settings that make it possible to collect data), the conversation could go like this: – John: I'm hungry and too lazy to cook. Google: Tables at the China Wok Restaurant are all reserved today, but I found this nice little place. . . • Demo: Google Brain (a futuristic scenario) http: //blog. outer-court. com/archive/2004_11_29_index. html 6 A. Frank-T. Sharon

People perspectives and privacy? (1) • Search is seen largely as a transient thing to many people. • I really don't stop to think much about the very personal things they look up. • I also have had no real experience with this information actively being recorded in a way they can use, and how it can be used by others, unlike with E-mail. 7 A. Frank-T. Sharon

People perspectives and privacy? (2) • Privacy is a fundamental concern with Web searches in general, and specifically with personalized search. • The ability for a search company to efficiently track and record personal search habits and tie them directly to someone’s identity has deep privacy implications. 8 A. Frank-T. Sharon

People perspectives and privacy? (3) • It’s not that personalized search will be necessarily harmful in terms of the privacy of personal information. • And perhaps users will choose to decrease privacy in order to increase efficiency or some other value. • But these issues need to be critically examined and debated in the public sphere. • This disregarding Scott Mc. Nealy's Law, Sun, 1999: "You have no privacy - get over it. " 9 A. Frank-T. Sharon

Googlism 10 A. Frank-T. Sharon

Personalized Google as an example • Question: “what kind of concrete privacy issues would you have with, say, a personalized Google? ” • Answer: numerous issues come to mind: – How do they track my search terms and links I click on? – How is this information matched with my Gmail, Calendar and other (future) Google services? – Do they sell this data to marketers? to Choice. Point? to private investigators? – What legal constraints exist to prevent law enforcement from getting this data? do they need a warrant? a subpoena? 11 A. Frank-T. Sharon

Google and privacy policy • "We do not rent or sell your personally identifying information to other companies or individuals, unless we have your consent. “ www. google. com/privacy. html • An interesting exception in their privacy policy though: "We provide such information to trusted businesses or persons for the sole purpose of processing personally identifying information on our behalf. When this is done, it is subject to agreements that oblige those parties to process such information only on our instructions and in compliance with this Privacy Policy and appropriate confidentiality and security measures. " 12 A. Frank-T. Sharon

"Hey, isn't this a little spooky? You know who I am!" 13 A. Frank-T. Sharon

Over-simplified Personalization? ! 14 A. Frank-T. Sharon

Over-simplified Personalization (1) • Over-simplified personalization stinks. • Human life is not as straight-forward as a variety of pastime checkboxes. • I might be interested in a restaurant near my home town, or I might be researching a restaurant for a friend living 100 miles away, and I'm still the same person I was before. • I might be looking to buy a plastic computer mouse today, and a living pet mouse tomorrow. 15 A. Frank-T. Sharon

Over-simplified Personalization (2) • Talking about mice, I might be watching Walt Disney's animated Basil of Baker Street, the Great Mouse Detective, and the same day cook something tasty which contains basil. • Much as we'd like to wish there would be no ambiguities in life (my, how boringly simple that would be). . . fact is, ambiguity's all around us and about to stay, too. • And what do we do to tackle it? Exactly: we provide context. Whereas the recipient, if not enough context is given, asks for more context. 16 A. Frank-T. Sharon

Over-simplified Personalization (3) • Google returning 2 million pages for your search is its way of asking for a refined search. It's like a friend saying, "Huh? What are you talking about? • Because if you type into Google "basil“, that just can't work. • Even if you would walk up to your best friend and say "basil", the friend wouldn't have the least clue what it is you are talking about. • Only if a context is established – say, for the "Basil" example, you are cooking with your friend, or you are watching a Disney movie with your friend, or you are referring to "basil recipe", or "basil the mouse detective" – is there a chance for this to make any sense. 17 A. Frank-T. Sharon

Over-simplified Personalization (4) 18 • The same happens in Google; you provide context. Since you can't watch a movie with Google or cook with Google, you choose the abstract level of human language (something we trained over thousands of years and upon which we build our culture, and certainly nothing to suddenly disregard now that computers start to try to help out). • It simply does not matter how well Google knows you, it will never be able to mind-read. Just like your best friend is not. And you get along good with your best friend, don't you? A. Frank-T. Sharon

Yahoo! Mindset (1) 19 A. Frank-T. Sharon

Yahoo! Mindset (2) 20 A. Frank-T. Sharon

Yahoo! Mindset (3) 21 A. Frank-T. Sharon

Yahoo! Mindset (4) 22 A. Frank-T. Sharon

Personalization and relevance rank 23 A. Frank-T. Sharon

Personalization improves relevance rank! (1) 24 • It's hard to believe that the generic relevance rank is the best it can get. When the generic relevance rank is computed, it makes a lot of assumptions, averages against a general population, and deals with noisy and incomplete information. More information should allow the results to be more relevant. • Let's take a Amazon search example. Try a search for "cookbook". If you've bought a few cookbooks on Chinese cooking, wouldn't those search results be more useful if they had a few books related to Chinese cooking sprinkled in there? A. Frank-T. Sharon

Personalization improves relevance rank! (2) • Clearly, you wouldn't want to show only books on Chinese cooking in this case. But there's some balance where a modest reordering to favor a couple books on Chinese cooking might be helpful. • The goal is to help people find the information they need. If using different relevance ranks for different people helps them find the information they need, that's what has to be done. 25 A. Frank-T. Sharon

Personalization doesn’t improve relevance rank! 26 • If I search for "cookbook" at Amazon, and I previously bought Chinese cookbooks, then this could mean a few different things: - I want to make a gift to my sister, so I'm basically searching for her. - I want to specifically find different cookbooks because I know how to cook Chinese now. - I want to get more Chinese cookbooks. • But simply entering "Chinese Cookbook" or similar solves the ambiguity, as is trivial to do. Any other way is neither trivial nor always possible in the first place. A. Frank-T. Sharon

Is generic relevance rank particularly objective? 27 • The core of the debate comes down to the saying: "No! I want to see the whole thing as objective as possible. " • Now, if you do see generic relevance rank as objective, I can see how the idea of fiddling with it would be unappealing. After all, you'd be taking something that is clean and precise and mucking it up with personalization goo. • But generic relevance rank isn’t particularly objective. The programmers put a lot of assumptions and educated guesses into it, mostly to deal with incomplete, imperfect, and noisy information about importance and relevance. A. Frank-T. Sharon

Subjective vs. Objective Search Results • Consequently, if more information can make your search results more accurate and relevant to you, that information should be used. • Subjective, the opposite of objective, means "particular to a given person". Personalized search results would be particular to me. I’m not sure I’d like that. • Search results that are fed by my personal prejudices will cease to be objective, and become subjective. Also, they will cease to be observed by others, thus cease to be objective in this sense. 28 A. Frank-T. Sharon

So how to overcome the subjective? • Regarding objective, the "prejudices" part means that the bias has to be preconceived and not based on knowledge. • But, regardless, the broader point here is that the personalization needs to be based on understandable, observable rules to make people comfortable. • There a couple ways to deal with that: – always explain the personalization, so the process becomes more transparent and observable. – always allow people to see the unpersonalized results. The personalization might be the default, but the unpersonalized results would be lower on the page or a click away, not missing entirely. 29 A. Frank-T. Sharon

Google Personalized – more or less 30 A. Frank-T. Sharon

Amazon: How Recommendations Work 31 A. Frank-T. Sharon

Amazon: Improving Your Recommendations 32 A. Frank-T. Sharon

Why is Personalization hard? • • • 33 Works from noisy, sparse information. Deals with changing preferences. Needs to understand human emotions. Needs to be done in real-time for many users. Has to make good predictions. Needs to avoid pigeonholing. Ø Conclusion: It can't ever be perfect. Personalization is so hard that it's going to make mistakes. Probably a lot of mistakes. A. Frank-T. Sharon

Local/Locality changes (technical) 34 A. Frank-T. Sharon

Context ambiguities aren’t hard to solve • Actually, this is not a problem, it’s the lack of a problem – and yet the biggest issue personalization tries to solve - ambiguity. When I enter “mouse”, do I mean computer mouse, Mickey Mouse, or a pet mouse? • Well, why don’t I just tell the SE? I couldn’t even call my best friend (who knows more than I’d ever want to share with a SE) and tell him “mouse”, and expect him to know what I want! • Why indeed not just enter “computer mouse” or “Mickey Mouse"? Or “Buy Computer Mouse” or “Mickey Mouse Video"? • This is a general approach which, once learned, always works. And as opposed to many problems associated with personalization, it’s trivial. 35 A. Frank-T. Sharon

People don’t like to authenticate • For personalization to work, SEs would need to provide means for people to authenticate themselves at different places/times. • Most people rather take the shortest route to achieving their goal, which would mean that explicit authentication is an additional hassle. • In any case, this could be achieved by login to the site, smart cards, biometrics, RFID tags, or other means of explicit authentication. 36 A. Frank-T. Sharon

!? Oops, I forgot my password 37 A. Frank-T. Sharon

People don’t like to fill forms • For personalization configuration/preferences, SEs would need to provide various means for people to provide relevant information. • However, filling out varied forms that have many different types of fields/settings is quite tedious and distracting. • SEs need to aim for minimal-input interfaces. Should configure most things automatically, but let users modify it if/as needed. 38 A. Frank-T. Sharon

(Want to be a My. Yahoo! user? ! (1 39 A. Frank-T. Sharon

(Want to be a My. Yahoo! user? ! (2 40 A. Frank-T. Sharon

(Want to be a My. Yahoo! user? ! (3 41 A. Frank-T. Sharon

Aha, now we’ll get you started! (1) 42 A. Frank-T. Sharon

Aha, now we’ll get you started! (2) 43 A. Frank-T. Sharon

Aha, now we’ll get you started! (3) 44 A. Frank-T. Sharon

Google personalized homepage 45 A. Frank-T. Sharon

Findory Personalization Technology 46 A. Frank-T. Sharon

People switch locations • People switch places all the time. • I might access the SE from work, from home, from an Internet Café, or from the road. • If the SE thinks it knows the common town a person is in, it might be wrong more than half of the time: a person can work in another city than he lives in; he could be on the road; he could be flying, etc… 47 A. Frank-T. Sharon

People don’t always mean “here” • Some people argue that it would be silly that when you enter “restaurant”, it doesn’t return only the restaurants close to you (Locality). • But very often, people are looking for information on places away from us. • Especially when going to travel to another city the next day and want to google for information on this place. . . like looking for a hotel. • Figuring out where a user wants to eat pizza if the same user omits the location is not trivial. 48 A. Frank-T. Sharon

”Where” & “Remember this location” 49 A. Frank-T. Sharon

“Enter address” & “My Locations” 50 A. Frank-T. Sharon

People/Times change (fundamental) 51 A. Frank-T. Sharon

People don’t know what they want in advance • People don’t know in what kind of areas they will be looking for in the future. They might find a new recreation, say. • Though it’s fun to play around with SE’s personalization options, it could be incredibly wary to use this in real life. • Why? Because I’m no expert on SE ranking algorithms. The SE should stay responsible for providing the best results to me. I can adjust my queries when I know specifically what I want. • But I don’t know today what I’m looking for tomorrow, and what kind of settings would be suited best by that time. 52 A. Frank-T. Sharon

Wait, my recreation is really “indoors”? ! 53 A. Frank-T. Sharon

People change over time • Personalization is usually based on past user behavior: – Analyzing the topics most often searched for. – Watching the results most often clicked on. – Remembering a personal “black list” of sites the user would prefer to avoid. • So why isn’t this always good? Because I may not want today what I wanted yesterday. • Learning is always getting to know things you do not already know. Analyzing people’s past will only make a SE understand what they already know, and what they were looking for once. • A personalized SE would risk feeding everyone their own prejudices. 54 A. Frank-T. Sharon

Hey, I broke my leg and had to change my hobby…I’m into music!!! 55 A. Frank-T. Sharon

People don’t like everything about something 56 • Let’s say my SE is trained to understand what things in life I like. That’s one approach to personalization: I tell my personal Google I like, say, movies. • But you know, I don’t like everything about movies. In fact, I probably dislike most movies. • In fact, because I like something in particular, my preferences in this area are much more refined; this makes me reject more here, not A. Frank-T. Sharon less.

Jane, get me off this crazy thing! 57 A. Frank-T. Sharon

People search for others • Some searches are for others. Say your friend or customer would like you to answer one of his questions. • You may not know much about the topic at hand, but you may understand better the power of SEs, so you go ahead and search for the other person. • Personalization would not only be useless in this case, it might be detrimental to the cause. The current search is not about the person that is performing the search. 58 A. Frank-T. Sharon

Your personalized google homepage … or 59 A. Frank-T. Sharon

Google classic homepage … 60 A. Frank-T. Sharon

People like to share • People like to share search queries. They may want to tell their friend: – “Enter ‘Apropos Tel-Aviv’ to find out the address of the cafe. ” • They may also want to share search URLs pointing to relevant results. • People may also want to universally discuss search results, and compare relevancy. • A personalized SE, all of a sudden, would inflict a barrier here. 61 A. Frank-T. Sharon

So, to use Personalization at all? • Does this all mean personalization is useless? Of course not. Personalization doesn't have to be perfect. It just has to be good enough. • If personalization helps people find what they need faster on average, it's a win. It doesn't have to be right all the time. Ø Personalized search will make mistakes - that's okay. It doesn't have to be perfect. It just has to be helpful. 62 A. Frank-T. Sharon

!? Personification or Penalization 63 A. Frank-T. Sharon