Today's personalization track is about checking the data and the evaluation before the models. One paper audits the public benchmark for recommending Hugging Face resources and finds knowledge graphs whose IDs do not match the items, plus item text that no method uses. Another uses Kuaishou's random-exposure data to measure how far category dashboards are skewed by the serving policy. On the modeling side, LMStaR fine-tunes a text embedding model to pick explanations for each user, and DualEvolve lets an agent improve both the recommender and its own way of running experiments.
Worth reading for anyone who evaluates recommenders, because the problem it uncovers is not in the models but in how the data files map onto each other.
Many teams decide which content categories to promote from the average watch ratio per category in their recommendation logs, but that dashboard is skewed by the serving policy.
When picking recommendation reasons for a user, earlier methods often fail to beat a simple baseline that only counts which statements appear most in the user's history.
When an agent improves a recommendation model on its own, existing frameworks usually judge each version only by its final score; nobody scores how well the experiment itself was run. DualEvolve scores the experimental process too, and beats the frameworks that look only at scores. It has two loops. The inner loop edits model and training code as usual. In the outer loop, a fixed LLM reads the round's experiment logs and scores them against a checklist, for example whether each change states a clear hypothesis, has a comparable control, and separates changes that did not help from runs that crashed. The agent then revises the rules it follows based on that score, and the next round runs under the new rules. The experiments pair four datasets with four recommendation models and track four ranking metrics for each pair; in all 64 of these comparisons, its test results beat OpenEvolve and DeepEvolve.
Different users often share mobility routines, but play them out at different places. DualClusterNet, a next-place recommender, learns routine and place separately. The model holds two sets of reusable vectors shared by all users: one describes the order in which check-ins happen, the other describes geographic space. The model encodes the user's check-ins so far into one vector and uses it to weight every vector in each set. The two weighted sums, one for routine and one for space, are multiplied element by element to form the current mobility preference, and a separate user-only branch keeps individual differences. On the New York and Tokyo check-in datasets, its top-1 prediction is the most accurate, but the lead is small, other methods do better when looking at the top few, and only one random seed was run.
Finding datasets is hard for researchers because, in scholarly knowledge graphs, datasets are often barely linked to papers and authors, and many lack text descriptions. MM-SAN recommends relevant datasets for a paper. It first finds, in the text of papers and datasets, mentions that match knowledge-base entries, and uses topic modeling to group the text into topics. These entities and topics become new nodes, so papers and datasets that mention the same entity or share a topic get connected. Each paper and dataset is represented from both graph structure and text, and the structure fills in when text is missing. In the hardest setting, where test nodes were never seen in training and only 25% of nodes have text, its recall is 41% higher than the strongest graph method.
WSDM 2027