Amancio Bouza holding the golden brain trophy for his proposal using collaborative filtering to create a win-win-win, build trust for clients and engage clients with UBS

From my experience of the people interacting with the Web 2.0, I built the following user model:

• like to share experiences and generate feedback
• like to generate content and to contribute
• want to be part of something bigger
• trust other users more then experts based on the Wisdom of Crowds assumption
• are intrinsic motivated
• are connected everywhere and every time
• do not honor guided help of experts or systems
• want do discover and explore

Based on this user model, I described the challenges of client advisory in general, showed how collaborative filtering meets the user model of the Web 2.0 and how collaborative filtering faces these challenges to build trust between client and adviser, empower the client to explore new possibilities, create adequate personalized product and service bundles, etc. . In additon, I proposed an framework that extends the current advisory process with collaborative filtering.

Finally, my proposal on “Collaborative Filtering – A Driver to Enable Clients to Explore, Share Experience and Build Recommendations for Products & Services” has been announced by the UBS jury as one of the three winners.
Congratulations to the additional two winners:

• Martin Moser & Roger Signer: UBS Super Trader
• Roland Studer: Go beyond ebanking of today

In addition, it has been awarded with the golden brain trophy from Starmind, a platform where expert knowledge and solutions to specific problem are traded. My proposal has been evaluated by the UBS jury as “creates a win-win-win and trust for clients and engages client with UBS”.

Further information can be read on:

• Blog article on Starmind.com about the contest
• News from the Department of Informatics at the University of Zurich
• News from the DDIS, the Dynamic and Distributed Information Systems group
• News from s.e.a.l., the Software Evolution and Architecture Lab

The contest has been announced on the 12th Web monday event in Zurich
On of the other winner published his proposal on “Go beyond ebanking of today”

Downloads

• Contest about Web 2.0 @ UBS

User interface of the spotting application

Some research has already been done investigating Web-based social networks and its applicability for different tasks such as trust inferrencing with trust networks or collaborative filtering respectively recommender systems. My master student Reto Hodel applied social network analysis to a social network and used the metrics to predict item ratings.
For this purpose, he built a very nice looking Web-based location recommender system called spotting. In this application, people get many locations presented on a geographical map that is based on Google Maps. The meaning of the encoded locations is very intuitive. The size of the location determines the level of match between the person’s preferences and the location. The bigger the location is presented the more relevant it is.
In order to boost the process of social network developing, we decided to build uppon an existing social network. Thus, we can rely on a already existent social network that has developed the relations among people already. For this reason, we decided to use the Web-based social network of Facebook because it provides a useful API, the Facebook API. Instead of developing a Facebook app, we decided to develop a stand-alone application that uses the Facebook Connect provided by Facebook. Facebook Connect is a function that a person can use to login to a different Web page with his Facebook account information. With this function it is still possible to gather his public information from Facebook, such as personal information and friends.
The value respectively usefulness of a collaborative recommender system increases with the amount of people and ratings. Therefore, the theory of network externalities respectively network effects applies to such systems. Thus, attracting people to use and rate the application in the first place is crucial.
Besides the social network and API’s that are provided by Facebook and others, Facebook also provides the facility of publishing news on the person’s wall that is seen by his friends. We used this function to publish location ratings and attract his friends to use spotting too. Our experience showed, that even with relying on a existing and developed social network it needs a lot of marketing to attract people. People do not just use an application because their friends use it. In general, no general approach exists to overcome this problem. First attempts try to reduce the so-called cold-start problem that describes the situation where only few ratings exist. But this is just one part of the solution.

Social Network of spotting

Anyhow, we could get 139 people in 3 weeks. Their ratings were the basis of our analysis. We analyzed the metrics of trendsetter, cliques, friend chains and some others. Unfortunately, we could not show that one social network metric leads to more accurate rating predictions. In contrast, the location’s average rating has been shown to perform better or at least equal to the social network graph patterns that we have investigated. In addition, the average rating is very cheap to compute! We applied the Friedman test that gave us evidence that some significantly performance difference exist on the significance level of 5%. Therefore, we run the Wilcoxon rank test. On the given significance level, we could not determine one single approach that is better then all the other. Of cause, we applied the Bonferroni correction for the family-wise error. The predictions based on trendsetters have been shown to be worst.
But these results have to be taken with caution. Despite applying statistical tests, the data set it self has some major threats to validity. The social networks consists mainly of the personal social network of the master student. That means that the average rating consisted mainly of his friends’ ratings. This fact highly influenced the performance of all other social network patterns.
However, I think that the experiment shows that the simple concept of wisdom of crowds respectively average rating is a simple but effective approach to provide ratings. Further, the average rating has the highest item coverage since patterns such as cliques, trendsetters, etc. do not give that much information about that many items.

Error distribution of cliques

Abstract

Generating accurate recommendations for items, such as locations, movies or books, is challenging. Common Web-based recommender systems require information about the users’ past to generate suitable recommendations for them.
In this thesis we first present spotting.li, a location recommender system based on Facebook, which allows users to rate locations and generates recommendations inferred by their friends’ ratings. In doing so, we examine requirments to successfully implement such a system using the latest web technologies (i.e., Grails) and describe key elements of our approach. Our focus is put on performance and providing an easy-to-use interface incorporating Google Maps.
Furthermore, we analyse different recommendation approaches which leverage structural in- formation from a social network to predict ratings. In particular, we examine the use of social network patterns, such as cliques and trendsetters, as well as direct friends and two levels of indirect friends. We finally conduct an extensive evaluation of these approaches, based on real data collected during the time of the thesis.
To prove our findings, we test our dataset, based on 139 users, for statistical significance. We demonstrate that even a simple algorithm, such as the average rating, bares similar results to more elaborate algorithms.

Reto Hodel: “spotting – Realisation and Analysis of a Location Recommender System Based on Facebook“, ed. by Amancio Bouza and Harald C. Gall, University of Zurich, December 2009. (master thesis)

User interface of the spotting location recommender system

In this master thesis, he built a Web applications that uses the people’s Social Network at Facebook to generate the location recommendations. Based on the data gathered by the Web application we analyze what SNA metrics are suitable for generating such recommendations. We are still investigating this research questions. As soon we get som e more insights we plan to add them to the Web application such that people can benefit from better recommendations.

If you are interest in having a look into the Web application then please just check out the following link:
http://seal.ifi.uzh.ch/spotting

You may want to join the spotting group at Facebook. So far, we only provide locations from Zurich.

Proposed Framework of Human-provided Services by David Schall

Current implementations of SOA only focus on services implemented as software. But software services may lack some ability that humans do not. Recently, the Web 2.0 revolution demonstrats the power of people’s collaboration and collaborative creativity.
Vitalab, the distributed system group of the TU Wien, invented the so-called Human-provided Services (HpS). HpS perfectly fit into the SOA such that they may even combined with other services for service composition, orchestration and choreography. User need to formally describe their user profile and services they intend to provide. You think that this approach is nonsense? Some applications exist already. In some cases, people are still more effective than the most sophisticated systems. For instance, Amazon provides a crowd sourcing service. People are asked to categorize images as images with or without pornographic content. Current algorithm are far from the accuracy and efficency o human analysis of visual images. By the way, that’s the method how professional spamer cracked the CAPTCHA security mechanism. They told users to enter the CAPTCHA code in order to get some more pornographic content.

I appreciate the invention of HpS. Imagine SOA-based systems combining computational power with man power. The use of HpS in recommender systems is straight forward. Some systems like FilmTrust by Jennifer Golbeck may benefit since recommendations rely on trust inferencing in Social Networks. But some people, especially outliers of the Social Network are mostly people participating in other Web-based Social Networks. Using them as HpS would lead to a more complete base for providing recommendations. But that’s just potential use in recommender systems.
But one may think about combining content filtering with HpS as hybrid recommender system where people may integrate the semantic meaning content information and contextual awareness. Even collaborative filtering approaches may benefit from better recommendations, since people predict the interests of other people very easy and associate interesting products with the person.

The Research day starts at 16.15 at the Department of Informatics. My adviser Prof. Abraham Bernstein opens the session with his talk on “Dem Gehirn beim Denken zusehen – Wie die Informatik neue Welten erschliesst”. Afterwards, all people present their current work on posters. Everyone will be open for questions and discussions.

I’ll participate and will present two different posters. One poster is about a distributed collaborative recommender system approach. My second poster is about a , OMORE, a Firefox Add-on that enables cross-site recommendation for movies.

You may find additional informations about the upcoming Research Day. Here you get important information how to reach us.

Distributed Collaborative Recommender System

OMORE - Firefox Add-on for cross-site recommendations

Participants of the IRMLeS workshop

Our current work on a probabilistic approach to compute partial user preference similarities was accepted and published at the 1st International Workshop on Inductive Reasoning and Machine Learning for the Semantic Web (IRMLeS) at the 6th European Semantic Web Conference (ESWC) 2009. The paper is available online at CEUR-WS.org as volume 474. The presentation is available at the IRMLeS Web page.

The general idea is that people may share similar preferences only partially. For instance, a person may like Italian food like another person but not Chinese food. But the other person does like Chinese food. Traditional collaborative filtering computes global preference similarity and fail detect this relation. Our approach computes is able to compute partial preference similarities on the basis of hypothesized user preferences. The hypothesized user preferences are learned applying traditional machine learning algorithms. We could show, that our approach performs significantly better then traditional user-based collaborative filtering. Especially in cases where people have only few common rated items. The strength of our approach are the use of partial preference similarities and using hypothesized user preferences instead of item ratings that are always biased.

It was my first real presentation at a conference and it was a great success. I got very positive feedback on it. But I also noticed that a too fancy presentation may irritate some people. Well, I just had the new version of Keynote installed on my Mac and thus, I had to try out the new fancy features. This workshop was one of the most successful at the this year’s ESWC measured by number of participants. I also enjoyed the workshop dinner where I participated interesting discussion on artificial intelligence, data mining in practice, football and Shakespeare.

At the conference, I got in touch with some very interesting people. Especially at the very well organised poster session and after the conference dinner. Unfortunately, it was the last European Semantic Web Conference because the organizers decided to have the industry as main target. Thus, the abbreviation of ESWC stands now for the Extended Semantic Web Conference.

Abstract

Recommender systems play an important role in supporting people getting items they like. One type of recommender systems is user-based collaborative filtering. The fundamental assumption of user-based collaborative filtering is that people who share similar preferences for common items behave similar in the future. The similarity of user preferences is computed globally on common rated items such that partial preference similarities might be missed. Consequently, valuable ratings of partially similar users are ignored. Furthermore, two users may even have similar preferences but the set of common rated items is too small to infer preference similarity. We propose first, an approach that computes user preference similarities based on learned user preference models and second, we propose a method to compute partial user preference similarities based on partial user model similarities. For users with few common rated items, we show that user similarity based on preferences significantly outperforms user similarity based on common rated items.

Presentation

In the following, you can watch my paper presentation I gave at the IRMLeS workshop:

Downloads

We include the papers on this page to ensure timely dissemination on a non-commercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by the copyrights. These works may not be reposted without the explicit permission of the copyright holder.

• PDF, alternatively from the Zurich Open Repository and Archive, DDIS group or s.e.a.l. group.
• BibTex
• Presentation slides

Amancio Bouza at the poster presentation of SemTree

Our current work on a ontology-based decision tree algorithm to learn user preferences was accepted and published at 7th International Semantic Web Conference (ISWC) 2008 in Karlsruhe (Germany). The paper is available online at CEUR-WS.org as volume 401.

It was my first participation of a conference where I had my first opportunity to present our early work on a novel approach on inducing a decision tree based on a domain ontology. SemTree is a machine learning algorithm based on the concept of decision tree induction. The novelty about SemTree is the usage of a domain ontology to learn more accurate models. That has been shown in a preliminary evaluation on a movie dataset.
At the poster session I had great conversations and discussions about recommender systems and my approach. But the disadvantage of presenting a poster to the research community is that oneself cannot take a look after the other interesting posters and demos.

To my surprise, Karlsruhe is a very nice german city dominated by students – a typical students town. That shows the high frequency of small bars and restaurants. The castle looks pretty and the very large castle park is absolutely beautiful. I would have liked to enjoy this park in the summer and spend my leisure time with friends there.

Abstract

Recommender systems play an important role in supporting people when choosing items from an overwhelming huge number of choices. So far, no recommender system makes use of domain knowledge. We are modeling user preferences with a machine learning approach to recommend people items by predicting the item ratings. Specifically, we propose SemTree, an ontology-based decision tree learner, that uses a reasoner and an ontology to semantically generalize item features to improve the effectiveness of the decision tree built. We show that SemTree outperforms comparable approaches in recommending more accurate recommendations considering domain knowledge.

Presentation

In the following, you can see a preview of the SemTree poster titled “seMANtics IN TREEs” in allusion to the american TV series “Man in Trees”:

Poster presentation of SemTree at the ISWC 2008

Downloads

We include the papers on this page to ensure timely dissemination on a non-commercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by the copyrights. These works may not be reposted without the explicit permission of the copyright holder.

• PDF
• RDF
• BibTex