Towards Question-based Recommender Systems

• Earlier conversational approaches use item preferences, predefined facets, dialogues, or facet-value data, while Qrec assumes items are distinguishable from associated text.The introduction identifies historical conversational data and hierarchical facets or facet-value pairs as assumptions Qrec drops.Source: §1 · p. 1
• The method combines collaborative filtering from matrix factorization with content analysis to provide explicit discrimination over informative terms.The paper argues that user feedback supplies new observations that can address sparsity and cold-start problems.Source: §1 · p. 2
• Qrec asks users about descriptive item characteristics extracted from item descriptions and reviews rather than eliciting preferences over items or facets.Its question pool is built from informative terms, and answers inform online latent-factor updates and belief-based question selection.Source: §1 · pp. 1–2
• Qrec achieves the highest performance among state-of-the-art baselines and is effective for both user and item cold-start recommendations.Source: §1 · p. 2

• Interactive conversational recommenders extend content-based, collaborative-filtering, and hybrid systems through varied forms of user feedback.Source: §2 · p. 2
• Interactive collaborative-filtering work elicits ratings, choices, or set feedback; Qrec instead asks about descriptive item terms and targets top-N recommendation.Compared with Christakopoulou et al., Qrec updates user and item latent factors and adapts questions to changing user preferences.Source: §2 · p. 2
• PMMN asks about review-derived aspect-value pairs, whereas Qrec uses ratings through matrix factorization and selects descriptive-term questions using user belief and GBS.The cited work also trains per user, while Qrec's online question answering can work for cold-start users and items.Source: §2 · p. 2
• Dialogue-based recommendation methods use conversation or interaction data, which may require repeated interactions or mentioned items; Qrec does not require such data.Source: §2 · p. 3
• Learning-to-ask approaches include data-hungry reinforcement learning and conversation-trained question generation; Qrec instead relies on user ratings and does not require duet-training histories.Zou and Kanoulas [44] require question-answering history and purchase behavior from the same query for duet training.Source: §2 · p. 3

3.1 Latent Factor Recommendation

• Qrec initializes user and item factors from ratings, then models both ratings and question-derived online affinity through a shared latent-factor framework.Its generative process includes separate parameters for rating and online-affinity observations.Source: §3.1 · p. 3

3.1.1 Optimization.

• Offline training learns factors from historical ratings; online ALS updates the active user and item factors after each answer.Offline optimization sets γ to 0 because historical question data is unavailable; online updates use closed-form solutions.Source: §3.1.1 · p. 4

3.2 Question Learning

• A Dirichlet belief tracker updates item-preference estimates from the prior and accumulated answers, and GBS selects entities that split preference mass nearly in half.The initial prior is based on the rating-derived item ranking.Source: §3.2 · pp. 4–5

3.3 Question Asking

• The question pool uses entities extracted from item documents, and questions ask whether the target item is associated with a selected entity.Users can answer yes, no, or not sure; the described extraction filters short-phrases below probability 0.1.Source: §3.3 · p. 5

3.4 Question-based Recommender System

• Qrec runs an offline rating-based initialization, then iterates question selection, answer-based belief and factor updates, and final ranking by updated user-item inner products.Historical interactions are assumed available for initialization, although the paper says the system can work without them.Source: §3.4 · pp. 5–6

4.1.1 Dataset.

• The dataset uses Amazon item descriptions and reviews from Home and Kitchen and Pet Supplies, with users and items appearing in fewer than five transactions filtered out.Interactions were randomly split 80%/10%/10% for training, validation, and testing.Source: §4.1.1 · p. 6

4.1.2 Parameter Setting.

• Offline matrix-factorization training uses 100 iterations and λu = λv = λp = λq = 0.1; γ, K, and question count are selected in RQ1.Source: §4.1.2 · p. 6

4.1.3 Evaluation Metrics.

• Recommendation accuracy is evaluated with recall@5, AP@5, MRR, and NDCG, with NDCG calculated over the top 100 items.An item is relevant when the user reviewed and rated it in historical buying behavior.Source: §4.1.3 · p. 6

4.1.4 Baselines.

• The five baselines are PMF, NeuMF, QMF+Random, SBS, and PMMN, spanning static recommendation and interactive question-asking methods.PMF, QMF+Random, and SBS share Qrec's parameter settings; NeuMF and PMMN use settings reported in their papers, with tuning following those papers.Source: §4.1.4 · p. 6

4.1.5 Simulating Users.

• Simulated users answer yes when an entity occurs in the target item's documents and no when it does not, assuming perfect knowledge.Source: §4.1.5 · p. 6

4.1.6 Online User Study.

• A small crowdsourced study has participants inspect a target product before answering system questions and viewing grids of 16 ranked items.Participants could stop answering at any time and then answered exit questions.Source: §4.1.6 · p. 6

4.1.7 Research Questions.

• The experiments ask about parameter effects, performance against prior methods, cold-start recommendation, offline initialization, and user-study evidence.Figure 2 focuses on γ and K across the two product categories.Source: §4.1.7 · pp. 6–7, Figure 2 · p. 7

• Performance improves on all metrics as questions increase, but gains are already strong within 10 questions and add little beyond 10.Source: §4.2 · p. 7
• The best γ is 0.5 in both categories, and performance first rises then falls as γ increases.γ controls how much online user feedback is incorporated into the user and item latent factors.Source: §4.2 · p. 7
• Performance also rises and then falls with latent-factor dimension K, so the paper uses the optimal value K = 3.Source: §4.2 · p. 7

• Qrec achieves the highest results on all four metrics against QMF+Random, SBS, and PMMN in both product categories.The authors attribute its advantage over QMF+Random to question selection and its advantage over SBS to the offline prior.Source: §4.3 · p. 7
• Qrec's matrix-factorization model outperforms PMF on all four metrics, while Qrec beats NeuMF on all four metrics with fewer than five questions.The authors note that NeuMF's stacked neural network is harder to train and incurs higher computational cost.Source: §4.3 · p. 7
• The authors report that interactive recommendation can improve over static models as it continuously learns from users.Source: §4.3 · p. 7

• The evaluation uses Home and Kitchen cold-start user tuples and Pet Supplies cold-start item tuples because the other category-specific tuple counts are too small for reliable results.The reported shares are about 84% cold-start user tuples in Home and Kitchen and about 7% cold-start item tuples in Pet Supplies.Source: §4.4 · p. 8
• Qrec achieves high recall@5, AP@5, NDCG, and MRR for cold-start users and items, whereas PMF does not work well in these cases.Table 3 reports cold-start users for Home and Kitchen and cold-start items for Pet Supplies.Source: §4.4 · p. 8, Table 3 · p. 8

4.5 Contribution of Offline Initialization (RQ4)

• Qrec with offline initialization outperforms random-initialized Qrec in both Home and Kitchen and Pet Supplies.The advantage is especially evident when fewer than 10 questions have been asked.Source: §4.5 · p. 8, Table 4 · p. 9

4.6 Online User Study (RQ5)

• Users answered an average of 15 questions per target item, and 71.4% said they were willing to answer 10–20 questions.The median number answered per target item was 12.Source: §4.6 · p. 9
• 95.2% of users said the questions were easy to answer, while 81% rated the experience positively.The remaining ratings were 14.3% neutral and 4.7% negative.Source: §4.6 · p. 9
• Users gave correct answers 95% of the time, were unsure 3.5% of the time, and gave wrong answers 1.5% of the time.Source: §4.6 · p. 9
• System performance measured after 5, 10, 15, and 20 queries, or when users chose to stop, agreed with Qrec’s Table 2 results.Source: §4.6 · p. 9

5 CONCLUSIONS AND DISCUSSION

• Qrec directly queries users about automatically extracted entities in relevant documents, updates latent factors online from answers, and tracks user belief to select question sequences.Source: §5 · p. 9
• User-study effectiveness results agree with the Qrec results reported in Table 2.Source: Table 5 · p. 9
• Questions use entity presence or absence in target items, while semantic relationships not lexically present in item documents are ignored.The authors also note that their user study is small and that a larger in-situ study would be more informative.Source: §5 · p. 9