Situating Recommender Systems in Practice: Towards Inductive Learning and Incremental Updates

“With information systems becoming larger scale, recommendation systems are a topic of growing interest in machine learning research and industry. Even though progress on improving model design has been rapid in research, we argue that many advances fail to translate into practice because of two limiting assumptions. First, most approaches focus on a transductive learning setting which cannot handle unseen users or items and second, many existing methods are developed for static settings that cannot incorporate new data as it becomes available. We argue that these are largely impractical assumptions on real-world platforms where new user interactions happen in real time. In this survey paper, we formalize both concepts and contextualize recommender systems work from the last six years. We then discuss why and how future work should move towards inductive learning and incremental updates for recommendation model design and evaluation.”

“Definition 2.1 (Transductive learning). In transductive learning, the algorithm knows all item contexts Itest as well as user contexts Utest that will occur in the test set upfront. Put differently, the algorithm will not be tested on out-of-domain contexts.

In most recommendation settings, the sets Utest and Itest are implicitly defined at training time to be the same as the ones seen in training. Returning to our earlier example, matrix factorization is transductive because at training time, we implicitly assume that all test contexts will be the same as during training, Itest = Itrain and Utest = Utrain.

Definition 2.2 (Inductive learning). In inductive learning, the algorithm has no information about the test set at training time. Hence, the algorithm must be able to work on unseen user-item contexts during testing.

There are two hybrid scenarios that arise when either information about user contexts or information about item contexts in the test set is available for training.

Definition 2.2.1 (User-inductive learning). In user-inductive learning, the algorithm knows only the item contexts Itest that will occur in the test set at training time. This requires the recommender to be able to accommodate unseen user contexts during testing.

Definition 2.2.2 (Item-inductive learning). In item-inductive learning, the algorithm knows only the user contexts Utest that will occur in the test set at training time.

We note that complementary to the learning scenarios above is the so-called cold start problem [39, 50]. In cold start recommendation, the focus lies on model performance behaves when not enough information about users or items is available (either at test or training time). This is independent of whether the model supports transductive or inductive inference.”

“Open challenges. Even though feature-based approaches offer a possible way to make a model inductive, there are fundamental limitations of such approaches. The most obvious limitation is that features are not always available in practice or cannot be accessed to due privacy reasons. Furthermore, features may fall short in representing diverse and latent aspects of user preferences [31]. Deriving good features requires substantial domain expertise; in many domains it is not even clear how semantically meaningful features should look like, e.g., art. This brings important research questions for future work: Can we combine feature-based information with explicit embeddings and still maintain an inductive model? If so, what are the trade-offs and how can be optimize them? Moreover, future work should investigate these questions with a dynamic environment in mind, as explicit features are typically static and cannot be readily updated as implicit representations.

Aggregation-based methods circumvent the issues above, but suffer from the curse of the directionality. Current aggregation-based methods have to make a choice and can only represent either items through users or vice-versa [24, 52]. This is fundamentally due to the directionality of the aggregation in the forward-pass as one has to serve as the inputs for the other. The fundamental quest here is: Can we design models that can form cyclic aggregation relationships between users and items for a fully inductive model? Recently developed methods, e.g., subgraph-based strategies [78] and permutation equivariant operators [18], are promising directions and can serve as good starting points.

The fact that latent factors typically require a gradient back-propagation process poses fundamental challenges to inductive learning. Future work should explore broader model families, such as non-parametric and tree-based models, to enable inductive recommendations. A bigger related question is also how properties of the data (sparsity, domain size, etc.) impact algorithm performance [11].”