General factorization framework for context-aware recommendations

Context-aware recommendation algorithms focus on refining recommendations by considering additional information, available to the system. This topic has gained a lot of attention recently. Among others, several factorization methods were proposed to solve the problem, although most of them assume explicit feedback which strongly limits their real-world applicability. While these algorithms apply various loss functions and optimization strategies, the preference modeling under context is less explored due to the lack of tools allowing for easy experimentation with various models. As context dimensions are introduced beyond users and items, the space of possible preference models and the importance of proper modeling largely increases. In this paper we propose a General Factorization Framework (GFF), a single flexible algorithm that takes the preference model as an input and computes latent feature matrices for the input dimensions. GFF allows us to easily experiment with various linear models on any context-aware recommendation task, be it explicit or implicit feedback based. The scaling properties makes it usable under real life circumstances as well. We demonstrate the framework’s potential by exploring various preference models on a 4-dimensional context-aware problem with contexts that are available for almost any real life datasets. We show in our experiments – performed on five real life, implicit feedback datasets – that proper preference modelling significantly increases recommendation accuracy, and previously unused models outperform the traditional ones. Novel models in GFF also outperform state-of-the-art factorization algorithms. We also extend the method to be fully compliant to the Multidimensional Dataspace Model, one of the most extensive data models of context-enriched data. Extended GFF allows the seamless incorporation of information into the fac[truncated]

💡 Research Summary

The paper introduces the General Factorization Framework (GFF), a versatile algorithmic platform designed to bring context‑aware recommendation into real‑world settings where implicit feedback dominates. Traditional factorization approaches such as SVD, FM, or iT‑ALS usually assume explicit ratings and rely on a fixed preference model (either an N‑way dot product or a pairwise interaction sum). Consequently, they either cannot exploit abundant implicit signals (clicks, purchases, views) or they lack the flexibility to experiment with alternative ways of incorporating multiple contextual dimensions (time, location, device, weather, etc.). GFF addresses both shortcomings by decoupling three core components of factorization methods: the loss function, the optimization routine, and, most importantly, the preference model.

The framework accepts any linear preference model over the involved dimensions. A linear model here means that the predicted preference for a tuple (user, item, context₁,…,contextₙ) can be expressed as a weighted sum of dot products or element‑wise products of the corresponding latent feature vectors. This minimal restriction enables researchers to plug in classic N‑way dot products, pairwise interaction sums, hybrid combinations, or even custom weighted schemes without rewriting the learning algorithm. The latent feature matrices for each dimension are learned jointly via an Alternating Least Squares (ALS) procedure. Because ALS updates one dimension at a time while keeping the others fixed, the computational cost scales linearly with the number of observed interactions and with the number of latent factors, making GFF suitable for large‑scale industrial data.

A crucial innovation is the treatment of implicit feedback. The observed user‑item‑context events are encoded as binary entries (1 for observed, 0 otherwise) in an N‑dimensional tensor. GFF introduces a weight function W(i₁,…,i_N) that assigns a real value to every possible tuple. For observed tuples the weight w₁ can be set to 1 (or a confidence derived from dwell time, recency, etc.). For missing tuples a separate weight w₀ is factorized across dimensions: w₀(i₁,…,i_N)=∑_j μ(j)·v_j(i_j)+γ(j). This factorization allows the model to express that missing data are not uniformly negative; instead, their influence can depend on properties of the user, item, or context (e.g., popular items receive higher missing‑data weight). The flexibility of w₀ is especially valuable for one‑class collaborative filtering where only positive signals are reliable.

The authors evaluate GFF on five real‑world implicit datasets, each enriched with four dimensions: user, item, time, and location. They compare several concrete models within the GFF family: (1) a pure N‑way dot product, (2) a pairwise interaction sum, (3) a time‑aware weighted N‑way model, and (4) a hybrid that also incorporates item metadata (genre, brand). Across all datasets, the best GFF configurations outperform state‑of‑the‑art baselines such as BPR‑MF, iT‑ALS, and Factorization Machines. Improvements range from 5 % to 9 % in Hit‑Rate@10 and Recall@10, and up to 3 % in NDCG when metadata is added. The experiments demonstrate that careful preference modeling—particularly the ability to weight contextual dimensions differently—has a material impact on recommendation quality.

Beyond the basic SA‑MDM (single‑attribute per dimension) setting, the paper extends GFF to be fully compliant with the Multidimensional Dataspace Model (MDM). MDM allows each dimension to contain multiple attributes (e.g., an item dimension may include ID, genre, price, and brand simultaneously). The extended GFF maps each attribute to its own latent matrix and introduces cross‑attribute weighting terms, thereby enabling seamless integration of rich side information such as social network connections, session logs, or textual tags. Preliminary experiments with item metadata confirm that the extended framework can capture additional signal without sacrificing scalability.

In summary, GFF delivers a four‑fold contribution: (1) native support for implicit feedback through a factorized confidence scheme, (2) a plug‑and‑play linear preference model interface that encourages rapid experimentation, (3) scalability comparable to classic ALS‑based factorization, and (4) full compatibility with the expressive MDM data model. By unifying these capabilities, GFF opens a practical research path for developing more accurate, context‑rich recommender systems that can be deployed directly in production environments.