PaperlayerView exact revision on arXiv

Source-linked AI summary

Seamlessly Unifying Attributes and Items: Conversational Recommendation for Cold-Start Users

Shijun Li, Wenqiang Lei, Qingyun Wu, Xiangnan He, Peng Jiang, Tat-Seng Chua

arXiv:2005.12979v5cs.IRcs.LGcs.SIstat.ML

TL;DR

Static recommendation depends on user histories, limiting its use for cold-start users, while conversational systems can ask about attributes before recommending items. ConTS unifies attributes and items as arms in a Thompson Sampling framework, and experiments report stronger performance than ConUCB and EAR on success rate and average turns.

  • Problem

    Static recommendation relies on historical user behavior, while cold-start users lack those histories and need recommendations from limited interaction.

    Source: §1 · p. 2

  • Method

    ConTS uses contextual Thompson Sampling to choose among attribute and item arms in one space, jointly deciding what to ask or recommend and when.

    Source: §1 · p. 3, §4.1 · p. 8

  • Results

    ConTS outperforms strong baselines including EAR and ConUCB on success rate in nearly all turns and average turn for successful recommendation.

    Source: §6 · p. 26

  • Takeaways & Limitations

    The experiments validate ConTS across Yelp, LastFM, and Kuaishou settings, with ablations supporting the contribution of initialization, reward estimation, and exploration.

    Source: §6 · p. 26

    The robustness analysis varies only ConUCB’s attribute-choosing function; item-choosing functions are not tested because adaptation requires complex mathematics.

    Source: §5.4 · p. 20

Abstract

from arXiv · show

Static recommendation methods like collaborative filtering suffer from the inherent limitation of performing real-time personalization for cold-start users. Online recommendation, e.g., multi-armed bandit approach, addresses this limitation by interactively exploring user preference online and pursuing the exploration-exploitation (EE) trade-off. However, existing bandit-based methods model recommendation actions homogeneously. Specifically, they only consider the items as the arms, being incapable of handling the item attributes, which naturally provide interpretable information of user's current demands and can effectively filter out undesired items. In this work, we consider the conversational recommendation for cold-start users, where a system can both ask the attributes from and recommend items to a user interactively. This important scenario was studied in a recent work. However, it employs a hand-crafted function to decide when to ask attributes or make recommendations. Such separate modeling of attributes and items makes the effectiveness of the system highly rely on the choice of the hand-crafted function, thus introducing fragility to the system. To address this limitation, we seamlessly unify attributes and items in the same arm space and achieve their EE trade-offs automatically using the framework of Thompson Sampling. Our Conversational Thompson Sampling (ConTS) model holistically solves all questions in conversational recommendation by choosing the arm with the maximal reward to play. Extensive experiments on three benchmark datasets show that ConTS outperforms the state-of-the-art methods Conversational UCB (ConUCB) and Estimation-Action-Reflection model in both metrics of success rate and average number of conversation turns.

1 INTRODUCTION

Static recommendation lacks the interaction histories needed for cold-start users, while conversational recommendation can elicit attribute preferences before recommending items. ConTS unifies attribute asking and item recommending as arm selection under Thompson Sampling.

1 INTRODUCTION

  • Conversational recommendation can elicit attribute preferences to reduce the candidate-item pool and improve item scoring for cold-start users.Cold-start users may interact only briefly, and their preferences can change over time.Source: §1 · p. 2
  • Existing item-only bandit methods cannot ask about attributes, while ConUCB separates attribute and item arms and uses a hand-crafted rule to time questions.The timing of attribute asking depends on factors such as the current turn and user feedback.Source: §1 · pp. 2–3
  • ConTS models attributes and items as undifferentiated arms, making three policy decisions one reward-maximizing arm-selection problem while using Thompson Sampling for exploration-exploitation balance.The decisions are what items to recommend, what attributes to ask, and whether to ask or recommend.Source: §1 · p. 3
  • ConTS outperforms Conversational UCB (ConUCB) and Estimation–Action–Reflection (EAR) on success rate and average conversation turns.The reported comparison is for cold-start users.Source: §1 · p. 4

2 RELATED WORKS

Related work spans static recommendation, online bandits, and conversational systems, but existing approaches leave gaps in handling cold-start users and jointly choosing attributes and items.

2 RELATED WORKS

  • Static recommendation methods rely on historical user-item records, while bandit methods pursue an exploration-exploitation trade-off but can struggle with large item pools.The paper notes that bandit methods commonly require separate candidate-pool preprocessing.Source: §2 · p. 4
  • Conversational recommendation includes NLP-based dialogue and form-based interactions, but many systems target existing users; EAR estimates preferences from existing-user records.Form-based systems use predefined dialogue paths and options.Source: §2 · pp. 4–5
  • Christakopoulou et al. found Thompson Sampling strongest among compared bandits for cold-start recommendation, but their system recommends items without asking about attributes.The paper argues that attribute preferences can filter large item pools before item selection.Source: §2 · p. 5
  • ConUCB models item and attribute arms separately and uses a turn-based hand-crafted asking rule, whereas ConTS unifies them in one arm space for contextual Thompson Sampling.ConUCB asks attributes when its frequency function increases between consecutive turns.Source: §2 · p. 5

3 PRELIMINARY

The preliminaries define the cold-start conversational setting and introduce contextual Thompson Sampling as the framework for balancing exploration and exploitation.

3.1 Problem Setting

  • In multi-round conversational recommendation, the system asks about attributes or recommends items until acceptance or a turn limit, seeking successful recommendations in few turns for users without interaction histories.User feedback can indicate liking or disliking attributes, accepting recommendations, or rejecting them.Source: §3.1 · pp. 5–6

3.2 Thompson Sampling

  • A contextual bandit selects an arm given context and receives user reward; in recommendation, the unknown parameter represents user preference.The goal is to maximize accumulated reward over limited interaction turns.Source: §3.2 · p. 7
  • Thompson Sampling derives a posterior from observed interactions, samples a preference parameter, and chooses the arm with the highest sampled reward.Sampling and updating support the exploration-exploitation balance.Source: §3.2 · p. 7
  • After feedback, contextual Thompson Sampling updates its preference estimate and uncertainty, lowering scores for negatively rated arms and similar embeddings while discouraging repeated exploration of known arms.The model updates its parameters using the chosen arm’s embedding and reward.Source: §3.2 · p. 7

4 METHOD

ConTS adapts contextual Thompson Sampling to conversational recommendation by initializing a new-user preference estimate, sampling under uncertainty, and choosing between unified attribute and item arms.

4 METHOD

  • ConTS unifies attributes and items in one arm space so the three conversational policy questions become a single arm-choice problem.The model aims to maximize reward while balancing exploration and exploitation.Source: §4 · p. 8

4.1 Overview

  • The workflow samples a user embedding, selects an arm by its estimated reward, then uses user feedback to update candidate pools and the embedding distribution.An item arm leads to recommending top-k items; an attribute arm leads to asking about that attribute or several top attributes.Source: §4.1 · p. 8

4.3 Arm Choosing

  • For arm choosing, ConTS compares attributes and items using a unified reward function that captures their mutual promotion, reducing action selection to choosing an arm.The selected arm determines whether the system asks an attribute or recommends items.Source: §4.1 · p. 8

4.2 Initialization and Sampling

  • ConTS trains a factorization machine on existing-user records to embed items and attributes together, initializes a new user’s mean preference with the average existing-user embedding, and samples from the resulting distribution.The covariance represents uncertainty about the estimated preference and controls exploration.Source: §4.2 · pp. 8–9

Algorithm 1 Conversational Thompson Sampling (ConTS)

ConTS unifies attribute questions and item recommendations as arm selection under contextual Thompson Sampling. It updates its user model from feedback and reports ranking complexity linear in candidate items and attributes, plus a quadratic embedding term.

Algorithm 1 Conversational Thompson Sampling (ConTS)

  • ConTS reduces choosing what to ask, what to recommend, and when to do either to selecting the arm with maximal reward.An attribute arm prompts a question; an item arm triggers recommendations of the top k items.Source: Algorithm 1 Conversational Thompson… · pp. 10–11

4.4 Updating

  • Negative feedback removes an attribute or item from the current arm pool, while accepted attributes update preferences and filter out items lacking them.Acceptance of a recommended item ends the session successfully.Source: §4.4 · p. 11
  • ConTS adjusts its user-embedding distribution using feedback while treating the initial user embedding and known preferred attributes as bias terms.The known preferences contribute to reward estimation; the posterior mean and covariance determine the updated distribution.Source: §4.4 · pp. 11–12

4.5 Complexity Analysis

  • ConTS has complexity O(2*d*(n+k) + d^2), for ranking n candidate items and k attributes and calculating an embedding-matrix inverse.The paper states that this adds no higher-order complexity than standard static recommendation and contextual Thompson Sampling methods.Source: §4.5 · p. 12

5 EXPERIMENTS

The experiments assess ConTS across three datasets and conversational settings, with additional comparisons of its components and performance under varied conditions. The data are divided into training, testing, and validation sets for offline training, cold-start conversation simulation, and tuning.

5 EXPERIMENTS

  • The experiments compare representative conversational recommenders and ablated ConTS variants, and examine exploration, attribute-item unification, turn limits, cold-start degree, and case studies.They also compare Thompson Sampling with UCB in this setting.Source: §5 · p. 13

5.1 Datasets

  • Yelp uses enumerated attribute questions, LastFM uses binary questions, and the new Kuaishou dataset uses multi-attribute questions.For Kuaishou, each attribute-asking turn presents the top 12 attributes; the dataset comes from cold-start users’ video-click records.Source: §5.1 · p. 13

5.1.1 Dataset Partition.

  • Users are disjoint across training, testing, and validation sets, so test users have no interaction records in the training set.Training is used for offline embedding learning, testing for cold-start conversation simulation, and validation for parameter tuning.Source: §5.1.1 · p. 13

Algorithm 2 Partition of Datasets

The experimental setup partitions users for cold-start evaluation, compares conversational recommenders and ConTS ablations, and evaluates simulated conversations using success and turn-efficiency metrics. Existing-user records train embeddings offline, while new-user simulations support online evaluation and parameter tuning.

Algorithm 2 Partition of Datasets

  • Algorithm 2 partitions by interaction-record proportion: training contains 70% of all user-item interaction records, and testing and validation each contain 15%; users in both held-out sets are new relative to training.Source: Algorithm 2 Partition of Datasets · p. 14

5.2.1 Baselines.

  • Baselines include item-only Abs Greedy, EAR, and ConUCB, which models attributes and items as separate arms.The ConUCB evaluation uses its Maximal Confidence attribute policy; the paper notes other policies are infeasible at its corpus scale.Source: §5.2.1 · p. 15
  • Three ConTS variants remove user-embedding initialization, known-preference terms, or Thompson Sampling’s exploration step.ConTS-u_init uses a zero-vector mean; ConTS-P_u removes preferred attributes from reward estimation; ConTS-exp uses the distribution mean each turn.Source: §5.2.1 · pp. 15–16

5.2 Experimental Setting

  • A user simulator gives positive feedback only for the target item and its attributes, and terminates unsuccessful sessions at the maximum turn limit.The primary experiments use a 15-turn limit; user parameters persist and continue updating across that user’s conversations.Source: §5.2.2 · p. 16

5.2.3 Evaluation Metrics.

  • Success rate SR@t measures conversations successfully completed by turn t, while average turns measures session length, capped at the turn limit.Higher SR@t and lower average turns indicate better performance; paired one-sample t-tests assess significance.Source: §5.2.3 · p. 16

5.2.4 Implementation Details.

  • Offline training uses existing-user records to fit user, item, and attribute embeddings with an FM, while online evaluation interacts with new-user simulations.ConTS parameters are tuned on simulated validation conversations.Source: §5.2.4 · p. 16

5.3 Experiment Results

ConTS significantly outperforms the compared methods on final success rate and average turns across the three datasets. Ablations indicate that preference attributes, exploration, and user initialization each matter to its performance.

5.3 Experiment Results

  • ConTS significantly outperforms all compared models on SR@15 and average turns across datasets, with higher success rates in nearly all turns except Yelp’s first several.Table 4 reports SR@15 and average turns for Yelp, LastFM, and Kuaishou; Figure 3 plots success rate by turn.Source: §5.3 · p. 17, Table 4 · p. 17
  • ConTS-P_u significantly underperforms ConTS, consistent with known preferred attributes contributing to reward estimation.Attribute asking also eliminates candidate items that lack the user’s preferred attributes.Source: §5.3 · p. 17
  • ConTS outperforms both EAR and ConTS-exp, which lack its online exploration mechanism.EAR exploits preferences learned during training; ConTS-exp removes Thompson Sampling’s exploration step.Source: §5.3 · pp. 17–18
  • ConTS outperforms ConTS-u_init; on Yelp, their success-rate curves gradually converge as conversation feedback accumulates.The authors report that initialization benefits diminish as the system gains more preference information, a trend not obvious on LastFM and Kuaishou.Source: §5.3 · p. 18

ConTS

ConTS unifies attribute asking and item recommendation in one arm space, while the discussion examines how that contrasts with ConUCB’s separate modeling and hand-crafted policy. The reported comparisons show sensitivity to ConUCB’s asking function and improved performance when its attribute-selection policy uses modified FM.

ConUCB

  • Figure 3 plots SR@t differences by turn against dataset-specific baselines, with ConTS higher on nearly all turns except the first several on Yelp.The baselines are ConTS-u_init on Yelp, ConUCB on LastFM, and EAR on Kuaishou.Source: Figure 3 · p. 19
  • Binary questions on LastFM and Kuaishou reduce the candidate pool less than enumerated questions, leaving larger pools and less efficient updates.Source: ConUCB · p. 19

5.4 Discussion: Unifying Attributes and Items

  • ConTS unifies attributes and items in one arm space and uses a unified reward function to reduce asking and recommending decisions to choosing an arm.The paper describes this as contextual Thompson Sampling and says the unified modeling captures mutual promotion between items and attributes.Source: §5.4 · p. 19
  • ConUCB models attributes and items separately, and its turn-based asking heuristic lacks a universally suitable function across datasets or scenarios.The paper argues that choosing a function is difficult because different functions may work better in different settings.Source: §5.4 · p. 19
  • On Yelp, ConUCB’s performance fluctuates hugely across functions for deciding whether to ask attributes; ConTS learns its policy from user feedback.Table 5 reports SR@15 and average turns for the different functions, with statistical significance reported at p < 0.01.Source: §5.4 · p. 20, Table 5 · p. 20
  • Table 6 shows that replacing ConUCB’s attribute-selection policy with modified FM improves its performance by a large degree.The comparison is between Maximal Confidence and modified FM as ConUCB’s attribute-selection policy on Yelp.Source: §5.4 · p. 20, Table 6 · p. 20
  • The mutual-promotion test replaces only ConUCB’s attribute-selection function with modified FM; the item-selection function was not tested.The paper says adapting the item-selection function requires complex mathematics.Source: §5.4 · p. 20

5.5 Discussion: Thompson Sampling v.s. UCB

The discussion separates the benefit of unified attribute-item modeling from the choice of bandit algorithm, then tests ConTS under tighter conversation limits. ConTS remains ahead under those limits, with its advantage growing as conversations allow more turns.

5.5 Discussion: Thompson Sampling v.s. UCB

  • Seamless-UCB, which keeps the unified item-attribute strategy in a UCB framework, still outperforms ConUCB.Table 7 compares ConTS, ConUCB, and Seamless-UCB; its caption notes significance testing at p < 0.01.Source: §5.5 · p. 21, Table 7 · p. 21
  • In this setting, ConTS using Thompson Sampling performs better than Seamless-UCB, though the authors leave the broader TS-versus-UCB comparison open.The authors note prior work reporting stronger Thompson Sampling performance in many real-world scenarios, but do not treat the broader comparison as settled.Source: §5.5 · p. 21

5.6 Discussion: Effect of Max Conversation Turn

  • The tighter-cap experiments compare methods at 7 and 10 turns; results are reported at turn 7 because several baselines perform poorly during the first five turns.The authors say turn-7 results better reveal differences between methods.Source: §5.6 · pp. 21–22
  • With 7- and 10-turn caps, ConTS significantly outperforms the other models on SR@7 or SR@10 and AT.The reported comparisons are in Tables 8 and 9.Source: §5.6 · p. 22, Table 8 · p. 22, Table 9 · p. 22
  • ConTS’s performance gap over other methods grows with the conversation cap; the authors attribute this to initially random policies and better preference capture across turns.They report that the difference is less obvious at small caps and accumulated advantage grows with more turns.Source: §5.6 · p. 22

5.7 Discussion: Effects of Cold-Start Degrees

The cold-start-degree analysis finds that ConTS is strongest with very little user history, while EAR gains more as history increases. Case studies show that the methods’ conversation policies can suit different user-interest patterns.

5.7 Discussion: Effects of Cold-Start Degrees

  • When historical records are very sparse (θ < 0.1), ConTS does better than EAR on Average Turn and SR@15 across all three datasets; with more records, EAR eventually outperforms it by a large margin.The study varies cold-start degree by reallocating randomly deleted user-item records, then trains an offline FM model for initial user embeddings.Source: §5.7 · p. 23, Figure 4 · p. 24

5.8 Case Study

  • In one Kuaishou case, ConTS shifts from a disliked attribute to exploring sports and recommends once it captures a preference, while ConUCB keeps asking and EAR favors popular themes.The example illustrates ConTS’s exploration of a new user’s interest.Source: §5.8 · p. 23, Figure 5 · p. 25
  • When a user’s interest is difficult to infer, ConUCB is the only method to complete the Buddhism-video recommendation within five turns, as its rule-based policy keeps asking.EAR and ConTS instead recommend after negative feedback to their initial questions in this example.Source: §5.8 · p. 24, Figure 5 · p. 25
  • For a user who likes a recently released movie, EAR quickly finds the popular movie attribute and recommends within three turns; the authors say it can be more efficient when new users resemble existing users.ConTS and EAR also explore less-popular attributes such as cars or history in this case.Source: Figure 5 · p. 25, §5.8 · p. 26

6 CONCLUSION AND FUTURE WORKS

The conclusion presents ConTS as a unified contextual Thompson Sampling approach for attribute questions and item recommendations. Experiments support its performance across several settings, while future work targets response handling, problem complexity, and neural extensions.

6 CONCLUSION AND FUTURE WORKS

  • ConTS treats attributes and items as undifferentiated arms, choosing what to ask or recommend through a unified reward-estimation and arm-selection process.The paper frames the three conversation-policy questions as a single arm-choosing problem.Source: §6 · p. 26
  • Across Yelp, LastFM, and Kuaishou, ConTS outperforms strong baselines on success rate in nearly all turns and average turns for successful recommendation.The reported baselines include Abs Greedy, EAR, and ConUCB; ablations also indicate benefits from initialization, reward estimation, and exploration.Source: §6 · p. 26
  • Future work includes handling “don’t know” or “don’t care” responses and preference comparisons, exploring reward estimation and update strategies, and extending the framework to neural models.The paper describes these as directions for more complex settings and preference modeling.Source: §6 · p. 26

7 SUPPLEMENTARY

The supplement explains how contextual Thompson Sampling updates a user’s preference estimate and uncertainty from interaction feedback. Treating items and attributes as equivalent arms lets the same updates inform both asking and recommending.

7 SUPPLEMENTARY

  • ConTS represents each user’s preference over arms with a multidimensional Gaussian distribution and updates its covariance after receiving arm feedback.The described arm has embedding x_a(t), and feedback is denoted r.Source: §7 · pp. 26–27
  • As actions increase the matrix diagonal, covariance decreases, reducing uncertainty and exploration for arms already played.The supplement says subsequent samples for those arms move closer to the mean preference estimate.Source: §7 · p. 27
  • Positive feedback raises the next score for the played arm and similar-embedding arms, while negative feedback lowers it, supporting exploitation of past experience.The supplement describes this as the effect of updating the preference mean.Source: §7 · p. 27
  • Because items and attributes are equivalent arms, the model updates their expected rewards in the same way and can choose between asking and recommending.The supplement presents this shared update as enabling action selection according to different situations.Source: §7 · p. 27
Loading 2005.12979v5…