Róbert Pálovics

Temporal influence over the Last.fm social network

In a previous result, we showed that the influence of social contacts spreads information about new artists through the Last.fm social network. We successfully decomposed influence from effects of trends, global popularity, and homophily or shared environment of friends. In this paper, we present our new experiments that use a mathematically sound formula for defining and measuring the influence in the network. We provide new baseline and influence models and evaluation measures, both batch and online, for real-time recommendations with very strong temporal aspects. Our experiments are carried over the 2-year “scrobble” history of 70,000 Last.fm users. In our results, we formally define and distil the effect of social influence. In addition, we provide new models and evaluation measures for real-time recommendations with very strong temporal aspects.

RecSys Challenge 2016: Job Recommendations

The 2016 ACM Recommender Systems Challenge focused on the problem of job recommendations. Given a large dataset from XING that consisted of anonymized user profiles, job postings, and interactions between them, the participating teams had to predict postings that a user will interact with. The challenge ran for four months with 366 registered teams. 119 of those teams actively participated and submitted together 4,232 solutions yielding in an impressive neck-and-neck race that was decided within the last days of the challenge.

Location-aware online learning for top-k recommendation

We address the problem of recommending highly volatile items for users, both with potentially ambiguous location that may change in time. The three main ingredients of our method include (1) using online machine learning for the highly volatile items; (2) learning the personalized importance of hierarchical geolocation (for example, town, region, country, continent); finally (3) modeling temporal relevance by counting recent items with an exponential decay in recency.For (1), we consider a time-aware setting, where evaluation is cumbersome by traditional measures since we have different top recommendations at different times. We describe a time-aware framework based on individual item discounted gain. For (2), we observe that trends and geolocation turns out to be more important than personalized user preferences: user-item and content-item matrix factorization improves in combination with our geo-trend learning methods, but in itself, they are greatly inferior to our location based models. In fact, since our best performing methods are based on spatiotemporal data, they are applicable in the user cold start setting as well and perform even better than content based cold start methods. Finally for (3), we estimate the probability that the item will be viewed by its previous views to obtain a powerful model that combines item popularity and recency.To generate realistic data for measuring our new methods, we rely on Twitter messages with known GPS location and consider hashtags as items that we recommend the users to be included in their next message.

RecSys Challenge 2014: an ensemble of binary classifiers and matrix factorization

In this paper we give our solution to the RecSys Challenge 2014. In our ensemble we use (1) a mix of binary classification methods for predicting nonzero engagement, including logistic regression and SVM; (2) regression methods for directly predicting the engagement, including linear regression and gradient boosted trees; (3) matrix factorization and factorization machines over the user-movie matrix, by using user and movie features as side information. For most of the methods, we use the GraphLab Create implementation. Our current nDCG@10 achieves 0.874. We release our experiments as IPython Notebooks.

Text Classification Kernels for Quality Prediction over the C3 Data Set

We compare machine learning methods to predict quality aspects of the C3 dataset collected as a part of the Reconcile project. We give methods for automatically assessing the credibility, presentation, knowledge, intention and completeness by extending the attributes in the C3 dataset by the page textual content. We use Gradient Boosted Trees and recommender methods over the evaluator, site, evaluation triplets and their metadata and combine with text classifiers. In our experiments best results can be reached by the theoretically justified normalized SVM kernel. The normalization can be derived by using the Fisher information matrix of the text content. As the main contribution, we describe the theory of the Fisher matrix and show that SVM may be particularly suitable for difficult text classification tasks.

Tutorial on Open Source Online Learning Recommenders

Recommender systems have to serve in online environments that can be non-stationary. Traditional recommender algorithms may periodically rebuild their models, but they cannot adjust to quick changes in trends caused by timely information. In contrast, online learning models can adopt to temporal effects, hence they may overcome the effect of concept drift. In our tutorial, we present open source systems capable of updating their models on the fly after each event: Apache Spark, Apache Flink and Alpenglow, a C++ based Python recommender framework. Participants of the tutorial will be able to experiment with all the three systems by using interactive Jupyter and Zeppelin Notebooks. Our final objective is to compare and then blend batch and online methods to build models providing high quality top-k recommendation in non-stationary environments.

Solving RecSys Challenge 2015 by Linear Models, Gradient Boosted Trees and Metric Optimization

The RecSys Challenge 2015 task requested prediction for items purchased in online web shop sessions. We describe our method that reached fifth place on the leaderboard by constructing a large number of item, session, and session-item features and using linear models and gradient boosted trees for learning. An important element of our method included optimization for the specific evaluation metric.