Tools that suggest charts for a table, such as a bar chart for category comparisons or a scatterplot for two numeric columns, increasingly adapt to the individual analyst. Current systems handle time in one of two ways. They either learn a single fixed taste per user offline, or they re-learn quickly inside each session and discard the result when the session ends. Neither fits how analysts behave. The same person hunts for anomalies with dense multi-variable scatterplots on Monday, builds aggregated bar charts for an executive deck on Wednesday, and goes back to anomaly hunting the following week. A system that treats this as "preference drift" and discounts old feedback will forget the exploration taste just before it is needed again. A system that treats it as a stable persona will push scatterplot matrices into the executive deck.
This note proposes a chart recommender that splits each analyst's preferences into two parts. The first is a persistent part that holds across all work, for example a dislike of pie charts or a preference for horizontal bars. The second is a small set of goal modes that recur, each with its own remembered preferences. The model infers from early in-session signals which mode is active and opens a new mode only when the evidence fits none of the old ones. The central empirical claim is testable: much of what looks like temporal preference drift in visualization use is switching among recurring goals. If that is true, remembering modes should beat both forgetting and averaging, and the gain should appear specifically when a user returns to an earlier goal.
Core Claim and What Is New
Two lines of chart-recommendation work bound this proposal. Personalized Visualization Recommendation (PVisRec) handles the fact that users analyze different private tables. It maps columns into shared statistical descriptors ("meta-features" such as entropy, skewness and cardinality) and maps charts into data-independent visual configurations. This makes cross-user learning possible, but each user gets one static embedding. Interactive Visualization Recommendation with Hier-SUCB treats recommendation as an online bandit, meaning a learner that trades off trying new charts against showing charts it expects the user to like. It picks a configuration first, then the attributes, and adds a learnable bias for the finding that users approve of an individually liked configuration paired with individually liked attributes only about 22% of the time. It adapts within a session and keeps no memory across sessions. Both papers motivate this work: PVisRec supplies the representation, and Hier-SUCB supplies the action decomposition and the evidence that online adaptation matters.
The treatment of history in this proposal differs from both. We score a candidate chart with configuration $c$ and attribute set $a$ for user $u$ in goal mode $z$ as
$$ s_{u,z}(c,a) = x(c,a)^\top(\theta_u + \phi_{u,z}) + b(c,a), $$
The terms are as follows. $x(c,a)$ is the meta-feature representation of the chart on the current table, following PVisRec. $\theta_u$ is the user's persistent preference vector. $\phi_{u,z}$ is the correction specific to mode $z$. $b(c,a)$ is Hier-SUCB's combination bias. Feedback updates the active mode's correction, and updates the persistent vector only when a signal holds consistently across modes. Each session begins with a belief over which of the user's existing modes is active, plus a "new mode" option weighted in the style of a Chinese restaurant process. That is a prior that makes opening a new cluster less likely as more clusters accumulate. Actual irreversible drift, in which the persistent vector itself changes, is handled by a separate change detector on persistent-level residuals. When that detector fires, it discounts only $\theta_u$.
This speaks directly to a tension in the literature. Memory systems such as MemoCRS, which stores timestamped (entity, attitude) triples, and Mem-PAL, which clusters past requirements into abstract "principles," assume that preferences stay continuous over time. Right Now, Wrong Then (NS-DPO) shows that uniform exponential discounting of old preference data rescues alignment after a preference flip, raising reward accuracy from about 50–55% to above 60–70%. However, its discount applies equally to every kind of preference. Our model uses discounting only where drift is actually detected and uses memory everywhere else. NS-DPO therefore provides the discounting mechanism we reuse for $\theta_u$, and it also defines the baseline we must beat.
Within a session, IACM-RL shows that marking replaced intents as stale, rather than deleting them, stops agents from sliding back to obsolete instructions. We borrow that idea for rejected charts. When a user pivots mid-session after an unexpected outlier, the charts they rejected are flagged stale for the old mode rather than being counted as negative evidence for the new one. This deliberately differs from DiscoverLLM, which models intent discovery as monotonic narrowing through an intent hierarchy without backtracking. Our modes can be left and re-entered.
Research Program
The first component is the model described above, built on Hier-SUCB's two-level action space so that exploration costs stay comparable to it. Mode inference uses signals that exist before much feedback arrives. These include dataframe operations, as in Lux, whose always-on engine already reads operation history and lightweight declared intent. They also include the chart types the user opens first and, where available, any natural-language request.
The second component is an evaluation environment. No benchmark in the file contains longitudinal, multi-session chart feedback with goal labels, so we build one in two layers. The first layer is a semi-synthetic simulator on the Plot.ly corpus used by PVisRec (17.4k users, 94k datasets). Each simulated user starts from a PVisRec-fitted embedding and receives two to four latent goal modes. Sessions follow scripted schedules that separate three regimes: recurrent switching, true irreversible drift, and stationarity. We will first check whether the corpus's per-user chart sequences can be ordered in time. If they can, we use them as a real-log test of whether recurring modes appear without being put there. The second layer is a small longitudinal study with analysts over several weeks, with sessions tagged afterwards as exploratory or presentational. Its purpose is to test whether the recurrent-mode regime actually occurs, not only whether the model works in simulation.
The third component is analysis. The model's claim is worthless if modes exist only in the simulator, so the main scientific output is an estimate, from real logs and the study, of how much session-to-session variation is explained by recurring modes and how much by monotone drift.
Baselines
The naive reference is a non-personalized ranker. We use the wide-and-deep model of Learning to Recommend Visualizations from Data, which reports nDCG@1 of 0.827 on a roughly 1,000-dataset Plot.ly subset against 0.366 for the best non-learned baseline. Any personalized method must clearly beat it on per-user held-out sessions, or personalization is not paying for itself.
The static personalized baseline is PVisRec's neural graph factorization. It reports +29.8% HIT@5 and +64.9% NDCG@5 over non-personalized baselines on the Plot.ly corpus. It represents the "stable persona" view, and we train it on all of a user's prior sessions.
The online baseline is Hier-SUCB restarted every session. Its paper reports surpassing the offline Neuro-PVR in hit rate at rank 1 within 80 interaction rounds. We found no reported absolute HR@1 on Lacuna for Hier-SUCB, so we re-run it ourselves.
The strongest competitor, and the one this proposal must beat, is Hier-SUCB warm-started across sessions from the previous session's parameters, with NS-DPO-style exponential discounting of past feedback. Its discount factor is tuned by grid search on validation users. This is exactly what a reviewer would build, and it covers both memory and drift with a single knob. We also run a language-model memory agent in the style of Mem-PAL. It summarizes past sessions into principles and re-ranks the same candidate charts, which isolates the value of structured modes compared with free-text memory.
All methods see the same candidate sets, feature representation, feedback sequence and hyperparameter tuning budget (the same number of validation configurations). The language-model agent gets a matched token budget.
Three ablations isolate the key ingredient. The first removes modes, leaving persistent preferences plus change detection. The second removes the persistent component, leaving modes only. The third gives the model oracle mode labels at session start, which bounds how much is lost to mode inference. Beating the discounted warm-start means lower cumulative regret and faster recovery of hit rate at rank 1 in the recurrent regime, with no loss in the drift and stationary regimes. Beating it only on simulated data does not count.
Why not just…?
Why not just warm-start Hier-SUCB and discount old feedback? This is the real threat, and the experiments are designed to settle it. In the true-drift and stationary regimes we expect a tie, because a well-tuned discount is close to optimal there, as NS-DPO's gains under flips and gradual drift suggest. The predicted separation appears only when a user returns to an earlier mode. Discounting will have erased that mode, while the mode memory recovers it within a few rounds. If real logs and the analyst study show that returns to earlier modes are rare, then this baseline is sufficient and we will say so. In that case the finding still tells the field that a single discount knob is enough for chart personalization.
Why not just give a language model the full history, or a memory summary? IACM-RL reports that baseline agents fall to 0% success on dialogues over 16k tokens in τ²-Bench, against 34.8% with explicit state tracking. MemoCRS reports up to 4.6× token savings from structured memory over raw logs. Both results argue against raw concatenation. Principle-style summaries such as Mem-PAL's encode continuity by construction and have no mechanism for retiring or reactivating a preference. The Mem-PAL-style baseline tests this directly on identical candidates.
Why not just ask the analyst which goal they are pursuing, or give them a mode toggle? This is partly right, and asking works when the goal is already formed. But AcCoRD finds frontier models scoring as low as 0.09 on preferences that only form after options are shown. Liu et al. find users anxious and inconsistent when their own concepts shift. Mid-session pivots triggered by an outlier are exactly the case where a toggle is not updated. The oracle-label ablation measures the value of perfect self-report. If inferred modes come close to it, asking adds friction without benefit. If they fall well short, a lightweight confirmation prompt becomes part of the design.
Proposed Experiments and Decisive Checks
On the simulator we report hit rate at rank 1, NDCG@5 and cumulative regret per round, broken down by regime. We also report recovery time, defined as the rounds needed to return to 90% of pre-switch hit rate. It is measured separately for switches into a previously seen mode and into a new one.
The decisive check is a crossover. The mode model should beat the discounted warm-start on returns to earlier modes and should tie it on novel drift. A uniform advantage would suggest a capacity artifact rather than the hypothesized mechanism. Varying the number of sessions between returns should show the discounted baseline's recovery time growing with the gap, while the mode model's stays flat. This follows because exponential discounting shrinks old evidence geometrically with elapsed time.
On time-ordered real logs, and in the analyst study, the key measurement is whether a held-out session's feedback is predicted better by the user's best-matching past mode than by the most recent sessions. That comparison is the empirical test of the claim that recurrent goals, not drift, dominate. In the study, the chart-level test is whether stale-flagging rejected charts after a pivot reduces repeated suggestions, compared with treating rejections as ordinary negative feedback.
References
- Personalized Visualization Recommendation
- Interactive Visualization Recommendation with Hier-SUCB
- MemoCRS: Memory-enhanced Sequential Conversational Recommender Systems with Large Language Models
- Mem-PAL: Towards Memory-based Personalized Dialogue Assistants for Long-term User-Agent Interaction
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
- IACM-RL: Intent-Aware Context Management and Reinforcement Learning for Complex Tool Invocation under Dynamic Intent Fluctuations
- DiscoverLLM: From Executing Intents to Discovering Them
- Lux: Always-on Visualization Recommendations for Exploratory Dataframe Workflows
- Learning to Recommend Visualizations from Data
- AcCoRD: Evaluating User-Agent Collaboration Under Realistic User Preference Dynamics
- Evaluating Rule-based Programming and ReinforcementLearning for Personalising an Intelligent System