PaperlayerView exact revision on arXiv

Source-linked AI summary

Interactive Path Reasoning on Graph for Conversational Recommendation

Wenqiang Lei, Gangyi Zhang, Xiangnan He, Yisong Miao, Xiang Wang, Liang Chen, Tat-Seng Chua

arXiv:2007.00194v1cs.IR

TL;DR

Traditional recommendation systems struggle to obtain dynamic, fine-grained preferences, and existing CRS methods use attribute feedback only implicitly. CPR addresses this gap by treating conversational recommendation as interactive path reasoning on a graph. SCPR outperforms existing CRS methods on Yelp and LastFM, with larger gains in larger attribute spaces.

  • Problem

    Traditional recommenders passively infer preferences from history, while existing CRS methods do not fully exploit explicit attribute feedback.

  • Method

    CPR models multi-round conversational recommendation as interactive path reasoning on a user-item-attribute graph, using feedback to navigate paths and restrict candidate attributes.

  • Results

    SCPR outperforms existing CRS methods on Yelp and LastFM across multi-round settings, especially when the attribute space is larger.

  • Takeaways & Limitations

    Graph-based path reasoning provides a more explicit and explainable way to use attribute preferences in multi-round conversational recommendation.

  • Takeaways & Limitations

    Evaluation relies on user simulators anchored to user-item interaction records, and CPR does not model adaptation after a user rejects a recommended item.

Abstract

from arXiv · show

Traditional recommendation systems estimate user preference on items from past interaction history, thus suffering from the limitations of obtaining fine-grained and dynamic user preference. Conversational recommendation system (CRS) brings revolutions to those limitations by enabling the system to directly ask users about their preferred attributes on items. However, existing CRS methods do not make full use of such advantage -- they only use the attribute feedback in rather implicit ways such as updating the latent user representation. In this paper, we propose Conversational Path Reasoning (CPR), a generic framework that models conversational recommendation as an interactive path reasoning problem on a graph. It walks through the attribute vertices by following user feedback, utilizing the user preferred attributes in an explicit way. By leveraging on the graph structure, CPR is able to prune off many irrelevant candidate attributes, leading to better chance of hitting user preferred attributes. To demonstrate how CPR works, we propose a simple yet effective instantiation named SCPR (Simple CPR). We perform empirical studies on the multi-round conversational recommendation scenario, the most realistic CRS setting so far that considers multiple rounds of asking attributes and recommending items. Through extensive experiments on two datasets Yelp and LastFM, we validate the effectiveness of our SCPR, which significantly outperforms the state-of-the-art CRS methods EAR (arXiv:2002.09102) and CRM (arXiv:1806.03277). In particular, we find that the more attributes there are, the more advantages our method can achieve.

1 INTRODUCTION

Traditional recommendation systems passively infer preferences, while existing CRS methods use explicit attribute feedback only implicitly. CPR instead models conversational recommendation as interactive graph path reasoning, and SCPR outperforms existing methods, especially with larger attribute spaces.

  • Traditional systems infer item preferences from historical actions, limiting access to dynamic, fine-grained, accurate, and explainable preferences.
  • Existing CRS methods can ask about preferred attributes, but typically encode attribute feedback implicitly through latent user representations.
  • CPR models conversational recommendation as interactive path reasoning over a user-item-attribute graph, using user-confirmed attributes explicitly.
  • Graph adjacency restricts candidate attributes to those connected to the current vertex, reducing irrelevant options and improving conversation coherence.
  • SCPR, a simple CPR implementation, outperforms state-of-the-art CRS methods on Yelp and LastFM across multi-round settings, particularly when the attribute space is larger.

2 RELATED WORK

Recommendation research has progressed from collaborative filtering and graph-based methods toward conversational systems that obtain explicit, dynamic attribute feedback. The paper argues that multi-turn conversational recommendation should use those attributes explicitly while exploiting graph structure for path reasoning and pruning.

  • Collaborative filtering models such as matrix factorization and factorization machines infer distributed user representations but cannot capture dynamic user preferences.
  • Graph-based recommendation methods use global connectivity or network embeddings to improve representations and exploit graph expressiveness.
  • Graph path-reasoning methods improve explainability by finding paths from users to target items, but remain static and require critical pruning because of high modeling complexity.
  • Conversational recommendation addresses dynamic preference and weak explainability by collecting explicit feedback, while existing multi-turn approaches do not use attributes explicitly.

3 MULTI-ROUND CONVERSATIONAL RECOMMENDATION SCENARIO

The multi-round conversational recommendation scenario lets a system repeatedly ask about attributes or recommend items, using user feedback to refine recommendations over conversational turns. It assumes accepted attributes strongly indicate preference and treats rejected attributes as items to remove from the candidate attribute set without modeling strong negative sentiment.

  • Scenario definition: MCR allows a CRS to ask attributes or recommend items multiple times, with each recommendation trial treated as one round.The system aims to obtain fine-grained preferences and provide accurate recommendations in the shortest conversational turns.
  • Interaction process: The conversation starts with a user-specified preferred attribute, after which the CRS selects either an attribute to ask or items to recommend.User feedback updates preferred attributes and candidate items during the session.
  • Assumptions: The scenario assumes that users express preferences clearly and that items containing all accepted attributes are sufficiently available.Consequently, the CRS prioritizes items containing every accepted attribute and ignores other candidates as a simplification.
  • Assumptions: Rejected attributes are treated uniformly as attributes the user does not care about, so they are removed from the candidate set without eliminating items containing them.The setting does not distinguish indifference from dislike because strong negative feedback is difficult to obtain in current data.
  • Decision problems: MCR addresses which items to recommend, which attributes to ask, and when to ask or recommend within a multi-turn interaction.These questions define the central decision problems of the scenario.

4 PROPOSED METHODS

CPR models conversational recommendation as interactive path reasoning over a user-item-attribute graph, explicitly navigating attribute feedback and using graph structure to constrain future choices. SCPR instantiates this framework with message-based scoring and a reinforcement-learning policy that decides whether to ask or recommend.

  • CPR framework: CPR represents users, items, and attributes as graph vertices and organizes their relations into a graph for conversational recommendation.Existing systems use the same information, whereas CPR exploits graph structure to conduct recommendation.
  • CPR framework: The active path records user-confirmed attributes chronologically, and CPR walks over attribute vertices to explicitly use conversational preference feedback.Previously visited attributes are excluded from further exploration.
  • Interactive process: Each CPR turn comprises reasoning, consultation, and transition: scoring items and attributes, choosing whether to ask or recommend, then updating the path and candidate sets after confirmation.The process repeats across conversation turns.
  • SCPR instantiation: SCPR scores items by propagating messages from user-preferred attributes and scores attributes from candidate-item messages, while its policy uses reinforcement learning to choose whether to ask or recommend.The policy therefore has two actions rather than selecting among all attributes directly, reducing its decision space.
  • Graph constraint: CPR restricts candidate attributes to those adjacent to the current attribute, excluding accepted and rejected attributes.This reduces the attribute search space and is intended to make dialogue more coherent than considering the whole attribute set.
  • SCPR instantiation: SCPR differs from EAR by using the adjacent-attribute constraint and graph message propagation for attribute scoring instead of relying on the policy network to score attributes.These choices leave the policy network with a smaller decision space.

5 EXPERIMENTS

The experiments evaluate SCPR against conversational recommendation baselines on Yelp and LastFM across multi-round settings, attribute granularities, and question types. SCPR outperforms the baselines, with larger advantages in the original large-attribute spaces, while its graph constraint and smaller action space support performance and explainability.

  • Experimental Setup: Experiments compare SCPR with Max Entropy, Abs Greedy, CRM, and EAR on LastFM and Yelp in the multi-round recommendation setting.The study considers binary and enumerated questions and both manually grouped and original attribute spaces.
  • Experimental Setup: The evaluation uses SR@t for cumulative success by turn and AT for average session turns, where higher SR@t and lower AT indicate better performance.Sessions are capped at T = 15 turns and recommendations contain the top k = 10 items.
  • RQ1: Performance Comparison: SCPR outperforms all baselines across the evaluated settings and shows larger advantages on LastFM* and Yelp*, which retain larger attribute spaces.The authors attribute this pattern to graph-based elimination of irrelevant candidate attributes and SCPR’s more dedicated reinforcement-learning policy.
  • RQ1: Performance Comparison: EAR may lead during the first few rounds but falls behind later, while Abs Greedy performs strongly early and plunges in further turns.The authors relate these patterns to differences in action spaces and whether methods exploit explicit attribute feedback for longer-term recommendation.
  • RQ2: Key Design: SCPR-v performs worse than SCPR in later turns, supporting the benefit of SCPR’s smaller action space; it still exceeds EAR because graph constraints narrow attribute choices.SCPR-v uses an EAR-style policy whose action space grows from 2 to |P| + 1.
  • RQ3: Explainability: SCPR produces an interpretable attribute path, whereas EAR’s less relevant attribute jumps lead to less coherent conversations and obscure recommendation reasons.A LastFM* example shows SCPR reaching a preferred artist through a short path from “metalcore” to “hardcore” to “post-hardcore.”

6 CONCLUSION AND FUTURE WORK

CPR addresses multi-round conversational recommendation by synchronizing graph-based path reasoning with conversation, while leaving several extensions for future work.

  • 6 CONCLUSION AND FUTURE WORK: CPR introduces graph-based path reasoning for multi-round conversational recommendation, using explicit attributes to improve explainability.The framework ranks items and attributes through message propagation and restricts transitions to adjacent attributes.
  • 6 CONCLUSION AND FUTURE WORK: Restricting transitions to adjacent attributes reduces the candidate space and makes the conversation more coherent.The policy network then decides only whether to ask or recommend, reducing its action space to 2.
  • 6 CONCLUSION AND FUTURE WORK: Future work includes handling rejected recommended items and developing more expressive attribute-scoring models.The current framework also leaves item and attribute embeddings unchanged during interactive training.

A DATASET STATISTICS

The dataset-statistics tables describe relation types for LastFM and Yelp, and separately report versions using original attributes to avoid complex feature engineering.

  • A DATASET STATISTICS: Table 4 lists relation types in LastFM and Yelp to clarify their dataset structures.The table is labeled as dataset statistics for the two datasets.
  • A DATASET STATISTICS: Table 5 reports statistics for LastFM* and Yelp* using original attributes to avoid complex feature engineering.The starred datasets are presented as separate dataset-statistics variants.

B DETAILS OF OFFLINE TRAINING IN THE REASONING STEP

The reasoning-step item-scoring component is trainable and is instantiated with the Factorization Machine model used in EAR for simplicity and reproducibility.

  • B DETAILS OF OFFLINE TRAINING IN THE REASONING STEP: The reasoning step uses a trainable item-scoring component instantiated as EAR’s Factorization Machine model.The paper specifies this instantiation to articulate the complete process and support reproducibility.

B.1 Training Objective

Offline training uses Factorization Machine-based scoring and pairwise objectives for both item and attribute prediction in multi-round conversational recommendation.

  • B.1 Training Objective: EAR embeds users, items, and attributes in one Factorization Machine for joint item and attribute prediction.The model is trained for the multi-round conversational recommendation scenario.
  • B.1 Training Objective: The scoring function models message propagation from users to items and from items to users.These correspond to the first and second terms of the scoring function.
  • B.1 Training Objective: Pairwise loss uses two negative-sample types, D1 and D2, tailored to multi-round conversational recommendation.D1 captures general user preference, while D2 incorporates confirmed attributes and a dynamically updated candidate-item set.
  • B.1 Training Objective: The dynamic candidate item set excludes interacted items and supports preference learning after attributes have been confirmed.The candidate set is obtained dynamically during the multi-round process.
  • B.1 Training Objective: Attribute prediction conditions the next attribute on the confirmed attribute set and is optimized with a pairwise loss.The prediction combines general preference for the attribute with its interaction with confirmed attributes.

B.2 Data Collection

The training datasets D1, D2, and D3 are collected through simulated multi-round conversational recommendation sessions grounded in observed user-item interactions. The procedure samples item instances under confirmed-attribute constraints and repeats simulations to improve dataset coverage.

  • Data collection: D1, D2, and D3 are accumulated through many simulated conversational recommendation sessions, with new instances appended at each interaction step.The simulations use a user simulator grounded in observed user-item interactions.
  • Data collection: Given a user and item attribute set Pv = {p0,p1,p2,p3,p4}, the procedure samples an item from the user’s non-interacted items.
  • Data collection: When confirmed attributes are Pu = {p0,p1,p2}, candidate items Vcand satisfy all attributes in Pu, and a negative D2 instance is sampled from the user’s non-interacted items in Vcand.The positive instance in these pairwise samples is the interacted item v for D1 and D2.
  • Data collection: All training-set user-item interactions ground the simulations, which are repeated for each interaction while trying every possible first user-informed attribute p0.This design is used to obtain high dataset coverage.

B.3 Training Details

The FM model is trained with a 64-dimensional embedding, SGD optimization, and L2 regularization. Separate learning rates are used for item prediction and attribute prediction.

  • Training details: The FM model uses an embedding size of 64, SGD optimization, and L2 regularization of 0.001.
  • Training details: The learning rates are 0.01 for item prediction and 0.001 for attribute prediction.
Loading 2007.00194v1…