Modeling user preferences has been mainly addressed by looking at users' interaction history with the different elements available in the system. Tailoring content to individual preferences based on historical data is the main goal of sequential recommendation. The nature of the problem, as well as the good performance observed across various domains, has motivated the use of the transformer architecture, which has proven effective in leveraging increasingly larger amounts of training data when accompanied by an increase in the number of model parameters. This scaling behavior has brought a great deal of attention, as it provides valuable guidance in the design and training of even larger models. Taking inspiration from the scaling laws observed in training large language models, we explore similar principles for sequential recommendation. Addressing scalability in this context requires special considerations as some particularities of the problem depart from the language modeling case. These particularities originate in the nature of the content catalogs, which are significantly larger than the vocabularies used for language and might change over time. In our case, we start from a well-known transformer-based model from the literature and make two crucial modifications. First, we pivot from the traditional representation of catalog items as trainable embeddings to representations computed with a trainable feature extractor, making the parameter count independent of the number of items in the catalog. Second, we propose a contrastive learning formulation that provides us with a better representation of the catalog diversity. We demonstrate that, under this setting, we can train our models effectively on increasingly larger datasets under a common experimental setup. We use the full Amazon Product Data dataset, which has only been partially explored in other studies, and reveal scaling behaviors similar to those found in language models. Compute-optimal training is possible but requires a careful analysis of the compute-performance trade-offs specific to the application. We also show that performance scaling translates to downstream tasks by fine-tuning larger pre-trained models on smaller task-specific domains. Our approach and findings provide a strategic roadmap for model training and deployment in real high-dimensional preference spaces, facilitating better training and inference efficiency. We hope this paper bridges the gap between the potential of transformers and the intrinsic complexities of high-dimensional sequential recommendation in real-world recommender systems. Code and models can be found at https://github.com/mercadolibre/srt.

When developing Java Script (JS) applications, the assessment and selection of JS packages have become challenging for developers due to the growing number of technology options available. Given a technology need, a common developers' strat-egy is to query Web repositories via search engines (e.g., NPM, Google) and shortlist candidate JS packages. However, these engines might return a long list of results. Furthermore, these results should be ranked according to the developer's criteria. To address these problems, we developed a recommender system called AIDT that assists developers in the package selection task. AIDT relies on meta-search and machine learning techniques to infer the relevant packages for a query. An initial evaluation of AIDT showed good search effectiveness. Recently, the emergence of ChatGPT has opened new opportunities for this kind of assistants, as reported by some experiments. Anyway, human developers should judge whether the recommendations (e.g., JS packages) of these tools are fit to purpose. In this paper, we report on a user study in which we used both AIDT and ChatGPT on a sample of JS-related queries, compared their results, and also validated them against developers' criteria and expectations for the task. Our initial findings show that ChatGPT is not yet on par with AIDT or even human efforts for the task at hand, but the model is flexible to be improved and furthermore, it can provide good arguments for its package choices.

When developing JavaScript (JS) applications, the assessment of JS packages has become a difficult and time-consuming task for developers, due to the growing number of technology options available. Given a technology need, a common developers’ strategy is to browse software repositories via search engines (e.g., NPM, Google) and identify candidate JS packages. However, these engines might return a long list of results, which often causes information overloading issues in the developer. Furthermore, the results should be ranked according to the developer’s criteria, but weighting the available criteria to choose a JS package is not straightforward. To address these problems, we propose a two-phase recommender system for assisting developers in retrieving and ranking JS packages in a semi-automated fashion. The first phase uses a meta-search technique for collecting JS packages that meet the developer’s needs. Based on criteria used by other projects on the Web, the second phase applies a machine learning technique to infer a ranking of relevant packages for the output of the first phase. We performed an initial evaluation of our approach with the NPM package repository and obtained satisfactory results in terms of both the accuracy of the retrieved packages and the quality of the ranking for the developers.