Personal, Private Movie Recommender System at the Semantic Web Challenge

I advised together with Gerald Reif the master thesis of Tobias Bannwart about a personal cross-site movie recommender system that is implemented as Firefox add-on. The add-on is known as OMORE and can be downloaded. We decided to bring OMORE to market maturity. For this, we had to rethink its architectural design and usability and beside from its advantages we came up with the following open challenges:

  1. Movie cross-references: It is not known what movie of one provider corresponds to what other movie page of another provider. Providers may be commercial pages like Amazon.com, review pages like RottenTomatoes.com or knowledge bases such as IMDb.com
  2. Retrieval of movie cross-references: No flexible search service exists to retrieve movie cross-references based on movie title and release year information.
  3. Maintenance of movie cross-references: A vast amount of potential movie cross-references exists that is difficult to gathered with a Web crawler approach. In addition, the set of movie pages increases fast.

We came up with the following solutions:

  1. Movie cross-references: A knowledge base is needed to persist movie cross-references. Concretely, for all movies the information of (1) what movie pages represent the movie as its content and (2) what movie pages represent the commercial product of a movie such as DVD, Blu-Ray, VHS or even Video-On-Demand (VoD). This semantical distinction has to be done, because a movie represented in VHS and Blu-Ray are not the same but show the same movie. Therefore, we decided to apply Semantic Web technologies to persist movie cross-references. We applied D2R that maps data from a relational database management system to RDF. RDF is the basic format to represent resources semantically. Our knowledge base of movie cross-references is called LiMo.
  2. Retrieval of movie cross-references: A search service has to be provided, that is able to provide even fuzzy search on the movie cross-references knowledge base. We reason for fuzzy search because the movie are presented quite heterogeneously among different Web pages. Movie titles may be misspelled, transformed or even extended in various way. Especially on online shops, we experienced that the movie titles are extended with information about many variants of special or collector’s edition and the type of medium the movie is provided. Instead of trying to extract the original title from the unpurified title, we decided to apply fuzzy search over movie titles and release year to retrieve movies. Our movie retrieval service can be accessed at MOLookup.
  3. Maintenance of movie cross-references: A Web crawler approach is not feasible due to the time latency and the need for resources. Thus, we decided to invent a collaborative approach. Whenever a user browses a new movie Web page that is not yet cross-referenced with LiMo, OMORE automatically uses the movie retrieval services MOLookup and provides the current URL of the new movie page. Then, this URL is cross-referenced to the retrieved movie. With that approach we automatically gather all the relevant movie cross-references with the user’s help. This way, normal users even contribute to the Semantic Web without knowing it.

With this new approach, we decided to participate in the this year’s Semantic Web Challenge that is co-located with the International Semantic Web Conference 2009 (ISWC) in Washington D.C. We were 16 participants that made it to the Semantic Web Challenge in Washington D.C. We presented our movie recommender system and its revised architecture besides the official Poster and Demonstrations session the main conference. Our secret weapon to attract many people to our stand was Swiss chocolate. And well, it worked out ;). The official time for the challenge presentation was 19:15-21:15. But people already showed up to our stand at half past 6 and kept coming by until 10 in the evening. One reason my be of course the chocolate ;), but also the viral marketing that people started that saw our challenge. Overall, people were really excited about our personal and private movie recommender system that even provides cross-site movie recommendations.
Despite the great success, we didn’t made it to finals. However, the challenge was a really nice experience and we had still have the great success having people excited about our OMORE.

Abstract

Online stores and Web portals bring information about a myriad of items such as books, CDs, restaurants or movies at the user’s fingertips. Although, the Web reduces the barrier to the information, the user is overwhelmed by the number of available items. Therefore, recommender systems aim to guide the user to relevant items. Current recommender systems store user ratings on the server side. This way the scope of the recommendations is limited to this server only. In addition, the user entrusts the operator of the server with valuable information about his preferences.
Thus, we introduce the private, personal movie recommender OMORE, which learns the user model based on the user’s movie ratings. To preserve privacy, OMORE is implemented as Firefox add-on which stores the user ratings and the learned user model locally at the client side. Although OMORE uses the features from the movie pages on the IMDb site, it is not restricted to IMDb only. To enable cross-referencing between various movie sites such as IMDb, Amazon.com, Blockbuster, Netflix, Jinni, or Rotten Tomatoes we introduce the movie cross-reference database LiMo which contributes to the Linked Data cloud.

Presentation

In the following, you can watch my presentation I prepared for the Semantic Web Challenge:

Poster

In the following, you can see a preview of the Semantic Web Challenge 2009 poster titled “OMORE”:
Poster of OMORE, a firefox plugin for online movie recommenations

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.

Personal Cross-Site Movie Recommender System Implemented as Mozilla Firefox Add-On

Online stores or Web page bring information about a myriad of items such as books, CDs, restaurants or movies at the user’s fingertips. Although, the Web reduces the barrier to the information, the user is overwhelmed by the number of available items. Therefore, online stores provide recommender systems that aim to guide the user to relevant items. However, recommender systems are generally limited to the Web page’s content and the explicit or implicit ratings provided by the users on the particular Web page. User are lazy when it comes to repeat providing rating information to recommender systems on other Web pages. That is a typical lock-in situation based on high transaction costs such that people are addicted to one or at least a limited number of Web pages.

People are required to have an account on a particular Web page before being provided with interesting recommendations. People may have concerns about providing explicit or implicit ratings on items that may expose some delicate details about their privacy. But many rather small online stores do not even provide recommendations.

Thus, we need a recommender system that (1) recognizes items over various Web pages as the same and remembers the ratings for those, (2) applies algorithms to provide recommendations and (3) smoothly integrates the rating and recommendation functionality directly in the Web site. We found the basic infrastructure for such a recommender system in the Firefox Add-on API and WEKA, a common data mining library.
We formulated these requirements as a master thesis. We were very happy to engage Tobias, an excellent master student.

The described recommender system implemented as Firefox Add-on can be downloaded at s.e.a.l. group site.

Abstract

Online stores and Web portals bring information about a myriad of items such as books, CDs, restaurants or movies at the user’s fingertips. Although, the Web reduces the barrier to the information, the user is overwhelmed by the number of available items. Therefore, recommender systems aim to guide the user to relevant items. Current recommender systems store user ratings on the server side. This way the scope of the recommendations is limited to this server only. In addition, the user entrusts the operator of the server with valuable information about his preferences.

In this thesis, we introduce our recommender system OMORE, a private, personal movie recommender, which learns the user model based on the user’s movie ratings. To preserve privacy, OMORE is implemented as a Mozilla Firefox add-on, which stores the user’s ratings and the learned user model locally at the client side. Although OMORE makes use of the movie features, which are provided by the different movie pages on the Amazon.com, Blockbuster, Netflix and Rotten Tomatoes.

Tobias Bannwart: “OMORE – Private, Personal Movie Recommendations implemented in a Mozilla Firefox Add-on“, ed. by Amancio Bouza, Gerald Reif and Harald C. Gall, University of Zurich, July 2009. (master thesis)

Downloads:

Probabilistic Partial User Model Similarity for Collaborative Filtering

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:

Presentation 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.

Evaluation of the Adoption of New Features in a Web-Based Social Network

Together with Marc Vontobel’s main advisor Gerald Reif, I successfully advised and supported Marc with his bachelor thesis. He was my first student I advised at all. He reenginered Purple Leaf – a party portal – and investigated the impact of social relations among people with respect to acceptance rate of new features on a web page. At the time he finished with his thesis, the community of Purple Leaf consisted of several thousands of people.
It seems that one rather accepts and adopts a new feature when a friend already adopted it. It seems that the inhibition threshold is much lower due to the fact, that people trust their friends and thus, trust the feature and it added value. Marc did a great job that could not be expected from a Bachelor student, starting from the software engineering skills to the data gathering and complex Social Network Analysis (SNA). But he was already prepared from our precedent seminar on “Trust and Recommendation in Social Networks”.

We were surprised as he delivered not only his thesis but also a complete picture book with fantastic analysis of differenc aspects of his party community. He discovered the core group of party people being on most parties, the latent semantics among different music styles and drinks.

Abstract

Purple Leaf is a social network which offers its member several possibilities to personalize its exclusive events by providing them unique online services. After the size of our platform suddenly increased from 300 initially invited guests to a multiple, we were obliged to completely revise the platform and enlarge our range of services. To embed these new services smoothly into the existing web presence, we fully restructured the application and changed the basis to a modern web framework. After that makeover, we designed five other services which we targeted to increase the customer loyalty and the entertainment value of our platform. Because new features are often not instantly accepted by existing users, we developed an integrated concept for boosting the acceptance of novel functionality. This concept is based on the technology acceptance model which was developed by Davis (1986). The model postulates that the actual use of a new feature is solely based on external factors. On the one hand, there are factors which influence the ‘perceived ease-of-use’ and on the other hand some that have impact on the ‘perceived usefulness’. In order to foster the perceived ease-of-use, we developed several usability concepts and tried to figure out how Web 2.0 features can help to simplify different processes. Beside the creation of intuitive user interfaces and plain procedures, we worked on an elaborated data and application structure which itself also contributed a big part to the simplicity of the new functionality. After we had embedded the services into our Internet portal, we started to analyze the acceptance of one new feature: ‘The most favored Guest’. This service allows every sign up member to define his personal list of favored guests for an upcoming event. Once the selected users are informed about their election, they, in turn, have the chance to define their own list. After a first round of selection, we tried to boost the personal acceptance of our members by providing specific incentives. Beside the active interventions into the process of adoption, we also analyzed a passive phenomenon: Does some kind of peer pressure exist within virtual cliques? If so, there might emerge some interesting changes in common marketing strategies which could narrow down the target audience to some single users of the network. In addition, we visualized some of the encountered situations and putted them together in an illustrated book as supplement to this paper.

Marc Vontobel: “Purple Leaf – Evaluation of the Adoption of New Features in a Web-Based Social Network”, ed. by Gerald Reif, Amancio Bouza, Harald C. Gall, University of Zurich, December 2008. (bachelorsthesis)