Skipforward is a distributed annotation system allowing users to enter and browse statements about items and their features. Items can be things such as movies or books; item features are the genre of a movie or the storytelling pace of a book. Whenever multiple users annotate the same item with a statement about the same feature, these individual statements get aggregated by the system. For aggregation, individual user statements are weighted according to a competence metric based on the constrained Pearson correlation, adapted for Skipforward data: A user gets assigned high competence with regard to the feature in question if, for other items and the same feature type, he had a similar opinion to the current user. Since the competence metric is dependent on the user currently viewing the data, the user's view of the data is completely personalized. In this paper, the personalization aspect as well as the item and expert recommender are presented.
Rating and recommendation web platforms have become important and ubiquitous nowadays. Typically, these platforms support collaborative ltering; users can rate items and are recommended items that people who liked the same items gave a high rating as well. This works ne for many cases and many domains; drawbacks are that in-depth explanations of recommendations cannot be given, and that the user has little control over the actual recommendation process. On the other hand, there is content-based ltering, which recommends items based on the features the user presumably likes. E.g., \We recommend song X since that song features prominent drums that you seem to like". Pandora.com is an example for such a system. This approach does not have the shortcomings of collaborative ltering outlined above; however, getting the content-based annotations needed for the recommendation process is costly, as this typically requires trusted experts.
Skipforward [
Skipforward's simple basic data model allows quite thorough annotation of items and provides rich metadata for recommender and other functionality. Conicting annotations do not break the system; a competence metric is used for weighting individual statements for aggregated views on the system's data. The competence metric is the foundation for much of Skipforward's recommendation functionality, which not only includes an item recommender that nds similar items or items tting some user-chosen feature pro le, but also an expert recommender, and annotation recommenders (functionality that helps annotating items). 2
For an overview of most Skipforward components, also see [
We have a number of requirements the top level ontology (coined Skipinions ) shall be able to handle.
The ontology should be able to represent user opinions of items such as books, movies, etc. { we solve this by providing an Item and a Feature class. For example, book features could be \Thriller (Genre)" or \Fast-paced writing style". Every Item can be associated to a Feature by using the Item's hasFeature property.
The facts databases of multiple users should be easy to merge. Fact databases can be just copied together using this approach. Smushing of items and features is done using the owl:sameAs predicate.
Provenance of statements needs to be tracked. We solve this by assigning an individual namespace to every user. Then, the URI of every instance created by this user has to use this namespace. This also ts nicely with Linked Open Data principles.
Plain text comments should be supported. Internationalized plain text comments can be added to Features.
It should be possible to explicitly dissent with an opinion of another user. This is implemented by the applicability property on every Feature. Applicability 1 means \this feature does not apply to this item", applicability +1 means
\this feature applies to this item", implementing the Open World Assumption. Additionally, any Feature can point to another Feature instance, implementing discussion threading.
Marking an opinion as uncertain should be possible. This is implemented by the confidence property on every Feature. Together with applicability, this forms a Dempster-Shafer-like approach.
The amount of noise seen by users should be kept minimal. The feature hierarchy as presented by the system is created by the owner of the respective namespace so arbitrary changes of the feature hierarchy are not possible. This is both a limitation and a feature of the system. It is limitating insofar as users cannot create new feature classes on the y. On the other hand, systems that implement this (i.e., most normal tagging systems) show that a lot of entropy enters the system otherwise. We try to keep this noise limited to the feature instance level where it can be handled in a coherent manner.
The basic top level ontology structure can be seen in Figure 1. Domain ontologies that subclass the Item and Feature classes as well as the hasFeature property are used for annotating actual items.
Currently, within Skipforward multiple domains are covered. Apart from
ontologies that model features of board games and a corresponding set of instances,
we mainly use DBTropes [
6 Note that the screenshot has been shortened for clarity. 2.3
The competence metric in Skipforward is based on user similarity with regard to
feature types. This is di erent from traditional content-based ltering as it takes
into account opinions of di erent users; it also di ers from standard collaborative
ltering which does not support multiple opinions concerning di erent feature
types per item. It is calculated with a modi ed variant of the constrained Pearson
correlation [
P i2Ixy
wxy;irx;iry;i i2Ixy wx;i(rx;i)2 P i2Ixy wy;i(ry;i)2 (1) For e cient calculation, an incremental algorithm has been implemented, only recalculating similarity values on changes, and only locally. simt(ux; uy) is used for weighting user statements for aggregated features. Aggregated features represent all user statements concerning one feature type and one item. They are used in the user interface and in recommenders. To compute the competence metric, the following algorithm is used.
Algorithm 2.1: CalculateAllCorrelations() for each i 2 I copyF eaturesT oCache(i) for each u 2 U
doInf erencing(u; i) for each tf 2 Tf calculateM ean(localuser; tf ) for each u 2 U calculateM ean(u; tf ) calculateCorrelation(localuser; u; tf )
To compute aggregated features per item and feature type according to the competence metric, another algorithm is used: For all feature types an item is annotated with, the aggregated feature for a speci c feature type is represented by a weighted sum of the individual features' applicability for this feature type. Applicability weight is the respective user's competence regarding the feature type.
In e ect, this means that for the aggregated feature, statements made by people who have been assigned a low competence value in uence the outcome less than statements made by people with a high competence value. 2.4
The item recommender shows items similar to the current item (Figure 3). It is based on a similarity metric comparing aggregated features assigned to the two items in question. For the standard item recommender user interface shown on each item page, this is a straightforward distance metric comparing feature applicability. Additionally, there is also an advanced recommender which lets users freely select and weight individual feature types, implementing recommendation channels. The annotation recommender is a utility for annotating items quickly. It shows a list of feature types that the current item has not yet been annotated with. Internally, for generating this list of feature types, item recommender output is reused. The annotations present for recommended items are compared with the annotations present for the current item. Any feature types for that a statement exists for the recommended item but not for the current item is presented to the user for quick annotation (Figure 4, right side). Therefore, annotating using the annotation recommender quickly improves the quality of results given by the item recommender.
Algorithm 2.2: GetRecFeatureTypes(curItem; recItems) for each i 2 recItems recF eatureT ypes:add(getF eatureT ypes(i)ngetF eatureT ypes(curItem)) 2.6
In Figure 4 (left side), the HTML interface representing the expert recommender
is shown. The expert recommender works on a per-feature type basis,
recommending users who expressed similar opinions concerning a selected Skipforward
feature type compared with the current user (simt(ux; uy) in Formula 1). Since
the current user agreed with the user Durham and disagreed with TV Tropes
(cf. Figure 2), TV Tropes was assigned a smaller weight for aggregation of the
feature type Battle Butler than Durham.
Skipforward is a unique amalgam of di erent technologies. Part of its
functionality can be found in other systems; for example, Revyu.com [
DBin [
In terms of recommendation functionality, Skipforward implements a
semantic hybrid ltering model. Similar approaches are discussed in [
A user study in the books domain will be carried out soon. A number of user interface improvements and additional statistics and recommender functionality will be added during that course. 4
This work has been supported by the German Federal Ministry of Education and Research (BMBF) in the context of the iGreen project (01IA08005A).