The paper presents our participation [5] at the ECML/PKDD 2011 - Discovery challenge for the task on the cold start problem. The challenge dataset was gathered from VideoLectures.Net web site that exploits a Recommender System (RS) to guide users during the access to its large multimedia repository of video lectures. Cold start concerns performance issues when new items and new users should be handled by a RS and it is commonly associated with pure collaborative lteringbased RSs. The proposed approach exploits the challenge data to predict the frequencies of pairs of cold items and old items and then the highest values are used to provide recommendations.
Recommender systems usually suggest items of interest to users by exploiting explicit and implicit feedbacks and preferences, usage patterns, and user or item attributes. Past behaviour is assumed to be useful to make reliable predictions, thus past data is used in the training of RSs to achieve accurate prediction models. A design challenge comes from the dynamism of real-world systems because new items and new users whose behaviour is unknown are continuously added into the system. As a consequence, recommendations may be negatively a ected by the well-known cold start problem.
Cold start is commonly associated with pure collaborative ltering-based
RSs. Particularly, item-based collaborative ltering techniques assume
that items are similar when they are similarly rated and therefore the
recommendations concern items with the highest correlations according
to the usage evidence. A straight drawback is that new items cannot be
recommended because there is not an adequate usage evidence.
Prediction involving cold items requires di erent approaches by
comparing the performance for the predictions about hot items. This may be
desirable due to other considerations such as novelty and serendipity.
Thus evaluating the system accuracy on cold items it may be wise to
consider that there is a trade-o with the entire system accuracy [
In addition, the dataset contains records about pairs of lectures viewed together (not necessarily consecutively) with at least two distinct cookieidenti ed browsers. This kind of data has a collaborative avour and it is actually the only information about the past behaviour. The user identi cation is missing, thus any user personalization is eliminated. User queries and feedbacks are also missing. 2
To overcome the cold start problem in the approaches based on
collaborative ltering, a common solution is to hybridize them with techniques
that do not su er from the same problem [
Data pre-processing step starts with obtaining an in-memory objectoriented representation of provided data.
The main output of this step is a set of 20 numeric values describing the similarities between lectures of each pair in the training set. The used features involve language, description, recording and publication ages, conference, authors and their a liations, and categories.
More details are reported in [
Model learning step allows to obtain a prediction model for the
frequency of a pair of lectures. The available data and the lightweight
goal determined the selection of a linear model for the learning
problem. Used features for di erent learned models are reported in [
Scale Problem
With the growth of the dataset, many recommendation algorithms are
either slowed down or require additional resources such as computation
power or memory. As such, it is often the case that algorithms trade
other properties, such as accuracy or coverage, for providing rapid results
for huge datasets [
The developed components allow to complete the recommendation task for the 5,704 lectures in almost 85 seconds on a notebook with an Intel Core 2 at 2.0 GHz as CPU and 2GB of RAM, i.e., each new recommendation about 30 cold items over the selected 1,122 ones is provided in almost 15 milliseconds. Reasonably, a production server allows to reduce further the response time for new recommendations and a cache speci cally devised for the recommendations allows to increase the throughput. 3
We have described the steps to achieve the submitted solution that
outperforms the random baseline at the ECML/PKDD 2011 - Discovery
challenge. The content-based hybrid approach allows to deal the cold
start problem. In addition it chances to provide also serendipitous
recommendations alongside classical ones [
Finally, the scalability performance is considered as a primary requirement and a lightweight solution is pursued. The preliminary performance for the notebook execution is quite promising and some future directions for improving latency and throughput are sketched.