Abstract New items, also called cold-start are introduced every day in the catalogs of numerous online systems. Due to absence previous preferences, recommending these items is difficult but important task for a recommender system. For this reason, item recommendation problem still represents an interesting research topic community. In work, we propose Neural Feature Combiner ( NFC ), novel deep learning, item-based approach recommendation. The model learns map content features into...

In this paper we provide a description of the methods used as team BanaNeverAlone for ACM RecSys Challenge 2020, organized by Twitter. The challenge addresses problem user engagement prediction: goal is to predict probability (Like, Reply, Retweet or with comment), based on series past interactions Twitter platform. Our proposed solution relies several features that extracted from original dataset, well consolidated models, such gradient boosting decision trees and neural networks. ensemble...

This paper focuses on recommender systems based item-item collaborative filtering (CF). Although research item-based methods is not new, current literature does provide any reliable insight how to estimate confidence of recommendations. The goal this fill gap, by investigating the conditions under which recommendations will succeed or fail for a specific user. We formalize CF problem as an eigenvalue problem, where estimated ratings are equivalent true (unknown) multiplied user-specific...

In this paper we provide an overview of the approach used as team PoliCloud8 for ACM RecSys Challenge 2019. The competition, organized by Trivago, focuses on problem session-based and context-aware accommodation recommendation in a travel domain. goal is to suggest suitable accommodations fitting needs traveller maximise chance redirect (click-out) booking site, relying explicit implicit user signals within session (clicks, search refinement, filter usage) detect users intent. Our proposes...

Cold-start is a very common and still open problem in the Recommender Systems literature. Since cold start items do not have any interaction, collaborative algorithms are applicable. One of main strategies to use pure or hybrid content-based approaches, which usually yield lower recommendation quality than ones. Some techniques optimize performance this type approaches been studied recent past. them called feature weighting, assigns every real value, weight, that estimates its importance....

This work presents a comprehensive study, from an industrial perspective, of the process between collection raw data, and generation next-item recommendation, in domain Video-on-Demand (VoD). Most research papers focus their efforts on analyzing recommender systems already-processed datasets, but they do not face same challenges that occur naturally industry, e.g., processing interactions logs to create datasets for testing. paper describes whole data including cleaning, processing, feature...

Most state-of-the-art top-N collaborative recommender systems work by learning embeddings to jointly represent users and items. Learned are considered be effective solve a variety of tasks. Among others, providing explaining recommendations. In this paper we question the reliability learned Matrix Factorization (MF). We empirically demonstrate that, simply changing initial values assigned latent factors, same MF method generates very different items users, highlight that effect is stronger...

In this paper we provide an overview of the approach used as team Trial&Error for ACM RecSys Challenge 2021. The competition, organized by Twitter, addresses problem predicting different categories user engagements (Like, Reply, Retweet and with Comment), given a dataset previous interactions on Twitter platform. Our proposed method relies efficiently leveraging massive amount data, crafting wide variety features designing lightweight solution. This results in significant reduction...

Hyper-parameter optimisation (HPO) is a fundamental task that must be performed in order to achieve the highest accuracy performance recommendation algorithm can provide. In recent past, with growth of dataset sizes, amount resources and time needed perform dramatically increased. Sampling data used during HPO procedure allows reducing required resources, but it impacts metric score. this paper, we study effects optimising hyper-parameters through random search, sampling users dataset. The...

This paper presents the solution designed by team "Boston Team Party" for ACM RecSys Challenge 2022. The competition was organized Dressipi and framed under session-based fashion recommendations domain. Particularly, task to predict purchased item at end of each anonymous session. Our proposed two-stage is effective, lightweight, scalable. First, it leverages expertise several strong recommendation models produce a pool candidate items. Then, Gradient-Boosting Decision Tree model aggregates...

Most state-of-the-art top-N collaborative recommender systems work by learning embeddings to jointly represent users and items. Learned are considered be effective solve a variety of tasks. Among others, providing explaining recommendations. In this paper we question the reliability learned Matrix Factorization (MF). We empirically demonstrate that, simply changing initial values assigned latent factors, same MF method generates very different items users, highlight that effect is stronger...

This paper describes the solution of our team PolimiRank for WSDM Cup 2022 on cross-market recommendation. The goal competition is to effectively exploit information extracted from different markets improve ranking accuracy recommendations two target markets. Our model consists in a multi-stage approach based combination data belonging In first stage, state-of-the-art recommenders are used predict scores user-item couples, which ensembled following 2 stages, employing simple linear and more...

Adding confidence estimates to predicted ratings has been shown positively influence the quality of recommendations provided by a recommender system. While over single point predictions and preferences widely studied in literature, limited effort put exploring benefits user-level indices. In this work we exploit recently introduced index, called eigenvalue order provide maximum for item-based systems. We firstly derive closed form solution calculate then propose new recommendation...