What sequential patterns emerge from anonymous single-session data?
This explores what a recommender can learn from a single anonymous visit — no login, no profile, just the order of clicks within one session — and whether that thin signal is enough to predict what someone wants next.
This explores what a recommender can learn from a single anonymous visit — no login, no profile, just the sequence of actions in one session. The surprising headline from the corpus is that this thin signal goes further than you'd expect. Can single sessions alone rival history-rich recommendation? adapts encoder-only transformers to session-based recommendation and finds that a single session, modeled well, can rival systems built on rich cross-session user history. The trick is treating the session like a sentence: mask the second-to-last action and learn to predict it, then slide that window across the visit. What emerges isn't a profile but a short-term intent trajectory — the within-session order of clicks turns out to carry most of the predictive weight.
The natural objection is that one anonymous session is sparse and easily ambiguous. The corpus answers this from the opposite direction: when individual histories are too thin to be useful, you can borrow structure from the crowd. Can cross-user behavior reveal news relations that individual histories miss? builds a global graph from aggregated, anonymous clicks across all users, surfacing article-to-article relationships that no single visitor's trail could reveal. So the field has two complementary moves on the same problem — model the session deeply on its own (sequential masking) or embed the sparse session inside a population-level behavioral graph (cross-user structure). The sequential pattern in a single session becomes legible because thousands of other anonymous sessions have already traced the paths.
There's a deeper pattern question lurking here: are short-session sequences just noise, or do they reflect something durable? Can language models discover what users actually want from activity logs? suggests the latter — that activity logs encode persistent 'interest journeys' lasting weeks, described in concrete phrases collaborative filtering never reaches. Read alongside the session work, this hints that the order of actions in even one anonymous visit may be a slice of a longer arc, not a self-contained whim. The challenge is that sequential systems must keep adapting as patterns drift; Can model isolation solve streaming recommendation better than replay? tackles exactly this with parameter isolation, preserving old behavioral patterns intact while adding capacity for emerging ones — the stability-vs-newness trade-off that any always-on session model lives inside.
The thing you might not expect to learn: 'anonymous' is doing less work than it sounds. Can LLMs predict demographics from social media usernames alone? shows models inferring gender, age, and politics from a username and sparse profile alone — meaning the sequential and behavioral signatures in supposedly anonymous data are richer than the missing login implies. Sequence order, click structure, and crowd-level patterns reconstruct a surprising amount of who someone is and what they'll do next, which is both why single-session recommendation works and why 'anonymous' deserves scare quotes.
Sources 5 notes
Sequential Masked Modeling adapts encoder-only transformers for session-based recommendation using penultimate-token masking and sliding-window augmentation. Across three datasets, this single-session approach consistently outperforms other single-session methods and rivals cross-session approaches with richer user history.
GLORY constructs a global news graph from aggregated user clicks to discover article relationships invisible in any single user's sparse history. This population-level behavioral structure enables recommendations even when direct textual or per-user similarity fails.
66% of users pursue valued interest journeys lasting over a month, described in specific phrases like 'designing hydroponic systems for small spaces.' LLM-powered journey discovery bridges the semantic gap that collaborative filtering cannot reach, operating at user-level granularity with persona-level precision.
DEGC uses per-task parameter isolation to handle streaming recommendation, providing explicit stability-plasticity trade-offs that experience replay and knowledge distillation methods cannot match. This approach preserves older patterns exactly while allowing new parameters to capture emerging preferences.
Evaluated on 1,384 survey participants and 48 synthetic accounts, web-browsing LLMs successfully predicted gender, age, and political orientation from X usernames and profiles alone. The models showed systematic gender and political biases specifically against low-activity accounts, relying on stereotype-driven defaults when content was sparse.