Paper deep dive
HyCoRec: Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation
Yongsen Zheng, Ruilin Xu, Ziliang Chen, Guohua Wang, Mingjie Qian, Jinghui Qin, Liang Lin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 87%
Last extracted: 7/21/2026, 5:04:25 AM
Summary
The paper introduces HyCoRec, a novel framework for Conversational Recommender Systems (CRS) designed to alleviate the Matthew effect (popularity bias) by modeling multi-aspect user preferences. It utilizes hypergraphs to capture complex relationships among items, entities, and words, alongside review and knowledge aspects, to generate diverse and accurate recommendations over time.
Entities (9)
Relation Signals (7)
HyCoRec → alleviates → Matthew Effect
confidence 95% · HyCoRec aims to alleviate the Matthew effect in conversational recommendation.
HyCoRec → uses → Hypergraph
confidence 90% · we propose a novel paradigm, Hypergraph-Enhanced Multi-Preference Learning... HyCoRec devotes to alleviate the Matthew effect by learning multi-aspect preferences... utilizing item-based hypergraph, entity-based hypergraph...
HyCoRec → models → Item-aspect Preference
confidence 85% · we derive the item-aspect preference P_i by leveraging the item-based hypergraph.
HyCoRec → models → Entity-aspect Preference
confidence 85% · we employ the entity-based hypergraph as a mechanism to learn entity-aspect preferences.
HyCoRec → models → Word-aspect Preference
confidence 85% · Derived from the word-based hypergraph, the word-aspect preference P_w can be formulated as...
DBpedia → usedby → Entity-aspect Preference
confidence 80% · we utilize the extensive DBpedia KG... to construct the entity-based hypergraph.
ConceptNet → usedby → Word-aspect Preference
confidence 80% · we build a word-based hypergraph using the word-oriented KG ConceptNet...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The Matthew effect is a notorious issue in Recommender Systems (RSs), \emph{i.e.}, the rich get richer and the poor get poorer, wherein popular items are overexposed while less popular ones are regularly ignored. Most methods examine Matthew effect in static or nearly-static recommendation scenarios. However, the Matthew effect will be increasingly amplified when the user interacts with the system over time. To address these issues, we propose a novel paradigm, Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation (HyCoRec), which aims to alleviate the Matthew effect in conversational recommendation. Concretely, HyCoRec devotes to alleviate the Matthew effect by learning multi-aspect preferences, \emph{i.e.}, item-, entity-, word-, review-, and knowledge-aspect preferences, to effectively generate responses in the conversational task and accurately predict items in the recommendation task when the user chats with the system over time. Extensive experiments conducted on two benchmarks validate that HyCoRec achieves new state-of-the-art performance and the superior of alleviating Matthew effect. Our code is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2607.17461v1
- Canonical: https://arxiv.org/abs/2607.17461v1
Trouble viewing inline? Open PDF directly →
Full Text
62,055 characters extracted from source content.
Expand or collapse full text
HyCoRec: Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation Yongsen Zheng1, Ruilin Xu1, Ziliang Chen2,5, Guohua Wang3*, Mingjie Qian1, Jinghui Qin4, Liang Lin1,2* 1Sun Yat-sen University, 2Peng Cheng Laboratory, 3South China Agricultural University 4Guangdong University of Technology, 5Jinan University z.yongsensmile, wangguohuagmial, scape1989@gmail.com xurlin5, qianmj7@mail2.sysu.edu.cn, c.ziliang@yahoo.com, linliang@ieee.org Abstract The Matthew effect is a notorious issue in Recommender Systems (RSs), i.e., the rich get richer and the poor get poorer, wherein popular items are overexposed while less popular ones are regularly ignored. Most methods examine Matthew effect in static or nearly-static recommendation scenarios. However, the Matthew effect will be increasingly amplified when the user interacts with the system over time. To address these issues, we propose a novel paradigm, Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation (HyCoRec), which aims to alleviate the Matthew effect in conversational recommendation. Concretely, HyCoRec devotes to alleviate the Matthew effect by learning multi-aspect preferences, i.e., item-, entity-, word-, review-, and knowledge-aspect preferences, to effectively generate responses in the conversational task and accurately predict items in the recommendation task when the user chats with the system over time. Extensive experiments conducted on two benchmarks validate that HyCoRec achieves new state-of-the-art performance and the superior of alleviating Matthew effect. Our code is available at https://github.com/zysensmile/HyCoRec. HyCoRec: Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation Yongsen Zheng1, Ruilin Xu1, Ziliang Chen2,5, Guohua Wang3*, Mingjie Qian1, Jinghui Qin4, Liang Lin1,2* 1Sun Yat-sen University, 2Peng Cheng Laboratory, 3South China Agricultural University 4Guangdong University of Technology, 5Jinan University z.yongsensmile, wangguohuagmial, scape1989@gmail.com xurlin5, qianmj7@mail2.sysu.edu.cn, c.ziliang@yahoo.com, linliang@ieee.org †footnotetext: *Corresponding author. 1 Introduction Conversational Recommender Systems (CRSs) engage in iterative conversations with users to provide personalized recommendations Qin et al. (2023); Li et al. (2023); Mishra et al. (2023), which have been widely adopted in various domains such as music recommendation Epure and Hennequin (2023) and online e-commerce Liu et al. (2023). Nevertheless, CRSs often face the prominent issue of Matthew effect Liu and Huang (2021), which can be described as “the rich get richer and the poor get poorer”. This phenomenon indicates that popular items/categories from past data receive more visibility in subsequent recommendations while less popular ones tend to be overlooked or ignored. Recently, many research efforts have focused on examining the Matthew effect in static or relatively-static offline recommendation scenarios Liu and Huang (2021); Anderson et al. (2020); Hansen et al. (2021). These offline studies strive to explore the potential causes behind the manifestation of the Matthew effect, and two key causes have been identified. One cause Anderson et al. (2020); Hansen et al. (2021); Liang et al. (2021); Zheng et al. (2021b) is that individuals with narrower and less diverse preferences exhibit a higher vulnerability to being trapped within the confines of the Matthew effect. Another cause Zheng et al. (2021c) is that the severe popularity bias where popular items consistently receive amplified exposure while less popular ones are underexposed. Although these methods have undoubtedly contributed valuable insights into the phenomenon of the Matthew effect, they directly overlook the adverse impact stemming from the dynamic user-system feedback loop. More recently, Gao et al. Gao et al. (2023) explore the Matthew effect in dynamic user-system interactions, but it lacks real-time user engagement through natural language conversations. Despite their effectiveness, most methods still suffer from two major limitations. 1) Interactive Schema. Many methods aim to mitigate Matthew effect in the static recommendation settings without considering the user-system feedback loop Zhang et al. (2021). In reality, the Matthew effect will progressively amplify as users dynamically interact with the system over time. Worse still, such amplification will inevitably lead to a series of notorious issues such as filter bubbles Steck (2018) and echo chamber Ge et al. (2020). Thus, it is crucial to consider the dynamic user-system interactions to alleviate Matthew effect. 2) Preference Learning. Prior studies Anderson et al. (2020); Hansen et al. (2021); Liang et al. (2021); Zheng et al. (2021b) show that the key to mitigating Matthew effect is to learn diverse user preferences. Thus, many methods leverage multiplex external Knowledge Graphs (KGs) to model multi-aspect preferences. But traditional KG edges are limited to linking only two vertices (i.e., factors), restricting preference learning to pairwise interactions. Instead, user relations exhibit intricate complexity, such as a user’s preference for a garment involves multiple factors like color, brand, style, and texture simultaneously. Hence, extending the number of vertices for learning diverse preferences is rather important. To address these issues, we propose a novel paradigm, Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation (HyCoRec), which consists of Hypergraph-Enhanced Multi-Preference Learning and Hypergraph-aware CRS. The former aims to model multi-aspect preferences, specifically targeting item-aspect, entity-aspect, word-aspect, review-aspect, and knowledge-aspect preferences. It addresses the Matthew effect in CRS by utilizing item-based hypergraph, entity-based hypergraph, word-based hypergraph, item reviews, and knowledge graphs to learn and derive these preferences. The latter focuses on leveraging these multi-aspect preferences as users interact with the system. Concretely, multi-aspect preferences are adopting to accurately predict the next utterances in the conversational task, and effectively make diverse item predictions in the recommendation task. By incorporating and utilizing these multi-aspect preferences, the system aims to provide precise and diverse recommendations that cater to the individual user’s preferences and needs for alleviating the Matthew effect as they continue to engage with the system. Empirically, extensive experimental results on two benchmarks show that HyCoRec outperforms all the compared baselines, and the superior of mitigating Matthew effect. Overall, our main contributions are included: • To the best of our knowledge, this is the first work to model multi-aspect user preferences, i.e., item-, entity-, word-, review-, knowledge-aspect preference, to alleviate Matthew effect in the CRS. • We proposed a novel end-to-end framework, HyCoRec, which adopts the multi-aspect preferences to effectively generate responses in the conversational task and accurately predict itms in the recommendation task. • Quantitative and qualitative experimental results on two CRS-based datasets exhibit superior performance of HyCoRec and the effectiveness of mitigating Matthew effect in the CRS. 2 Related Work 2.1 Conversational Recommender System Unlike traditional recomemnder system Zheng et al. (2021a); Jin et al. (2023); Zheng et al. (2024d, b, 2026a), Conversational Recommender System Qian et al. (2023); Zheng et al. (2024e, c, a, f, 2025b, 2025a, 2026b) aims to capture user preferences through dialogues and provide high-quality recommendations. Previous research on CRS can be broadly categorized into two main types: attribute-based CRS Deng et al. (2021a); Lei et al. (2020a, b); Ren et al. (2021); Xu et al. (2021) and generation-based CRS Chen et al. (2019); Deng et al. (2023); Li et al. (2022); Zhou et al. (2020a, 2022); Shang et al. (2023). Attribute-based CRS involves capturing user preferences by asking questions about item attributes and generating responses using pre-defined templates Lei et al. (2020a). But this strategy often neglects the importance of generating responses that resemble natural human language, which can negatively impact the user experience. On the other hand, generation-based CRS tackles this issue by utilizing the Seq2Seq architecture Vaswani et al. (2017a) to integrate both conversation and recommendation tasks to produce fluent and coherent human-like responses. Despite their effectiveness, they fail to model users’ diverse preferences since the user-item interactions data is rather sparse and limited. In contrast, our work aims to model multi-aspect preferences for exploring user diverse intricate relation patterns. 2.2 Matthew Effect in Recommendation Matthew effect is a notorious issue in RSs. Recently, Liu et al. Liu and Huang (2021) have substantiated the occurrence of the Matthew effect in YouTube’s recommendation system. Besides, Wang et al. Wang et al. (2019) undertook a rigorous quantitative analysis, offering valuable insights into the quantitative characteristics of the Matthew effect in recommender systems based on collaborative filtering. To alleviate Matthew effect, one common method is to consider recommendation diversity strongly advocated by researchers Anderson et al. (2020); Hansen et al. (2021); Liang et al. (2021); Zheng et al. (2021b), and another critical perspective is by removing popularity bias, a factor that has been identified as a catalyst for its amplification Zheng et al. (2021c). But these methods predominantly focus on investigating the Matthew effect in the static recommendation settings without considering the user-system feedback loop. Instead, our HyCoRec aims to alleviate Matthew effect considering dynamic user-system feedback loop. 3 HyCoRec Matthew effect is a notorious issue in the CRS, and it inevitably becomes intensified over time due to the existence of the dynamic user-system feedback loop. To tackle these challenges, we propose a novel paradigm, HyCoRec, which consists of Hypergraph-Enhanced Multi-Preference Learning and Hypergraph-Aware CRS. The overall pipeline of our HyCoRec is depicted in Fig.1. 3.1 Preliminaries 3.1.1 Conversational Recommendation Conversational recommendation is a personalized approach where the system engages in continuous dialogues with users to gain a deeper understanding of their preferences and deliver customized suggestions. This interactive method enables the system to collect additional insights into the user’s preferences, context, and requirements, resulting in more precise and relevant recommendations. CRSs are widely applied across diverse domains like e-commerce, music streaming, movie recommendations, and others, aiming to enrich user experience and satisfaction. 3.1.2 Hypergraph Hypergraphs demonstrate intricate configurations, capturing intricate relationships among numerous elements through hyperlinks. In our research, we depict user inclinations by constructing multi-grained hypergraphs, encompassing the item-based hypergraph item(t)G^(t)_ item, entity-based hypergraph entity(t)G^(t)_ entity, and the word-based hypergraph word(t)G^(t)_ word. Each hypergraph can be delineated as item(t)=(ℐ∗(t),ℋ∗(t),∗(t))G^(t)_ item=(I^(t)_*,H^(t)_*, N^(t)_*), comprising: (1) a node collection ℐ∗(t)I^(t)_*; (2) a hyperege ensemble ℋ∗(t)H^(t)_*; (3) a |ℐ∗(t)|×|ℋ∗(t)||I^(t)_*|×|H^(t)_*| adjacent matrix ∗(t) N^(t)_* signifying the weighted link between each node and hyperedge. 3.2 Hypergraph-Enhanced Multi-Preference Learning Extensive experiments by most existing methods Hussein et al. (2020); Liu et al. (2021); Nguyen et al. (2014) have consistently shown that users with restricted preferences are highly influenced by the Matthew effect. Thus, the key to alleviating such bad effect is to model the diverse user preferences. Along this line, we formulate the Hypergraph-Enhanced Multi-Preference Learning, including Multi-Hypergraph Construction and Multi-Preference Learning. 3.2.1 Multi-Hypergraph Construction Traditional KGs focus on pairwise interactions for preference learning, as edges connect only two vertices. However, user preferences often exhibit complex item relation patterns. To address this, we construct multiple hypergraphs (item-aspect, entity-aspect, and word-aspect), enabling connections between more than two vertices. Figure 1: Overview of our HyCoRec framework, which consists of Hypergraph-Enhanced Multi-Preference Learning and Hypergraph-Aware CRS. The former aims to dynamically learn multi-aspect user preferences, while the latter contains the conversation task to generate diverse responses and the recommendation task to predict target items. Item-based Hypergraph. Items directly reflect users’ genuine preferences. Users might prefer related items, such as products from the same brand or with similar features. Thus, establishing connections among similar or functionally similar items is crucial for exploring a diverse preferences. To do this, we first extract items from a session and treat them as vertices, forming a hyperedge. Then, All hyperedges associated with a user are connected through shared items to create the item-based hypergraph item(t)G^(t)_ item as: item(t) ^(t)_ item =(ℐi(t),ℋi(t),i(t)). =(I^(t)_i,H^(t)_i, N^(t)_i). (1) where ℐi(t)I^(t)_i means the item set extracted from the historical conversations, ℋi(t)H^(t)_i is the hyperedge set, and i(t)∈0,1|ℐi(t)|×|ℋi(t)| N^(t)_i∈\0,1\^|I^(t)_i|×|H^(t)_i| is the incidence matrix, which can be defined as: v,h(t)=1,ifv∈h0,ifv∉h N^(t)_v,h= \ aligned &1,&& if v∈ h\\ &0,&& if v∉ h\\ aligned . (2) The degree of a vertex v∈ℐi(t)v ^(t)_i is denoted as d(v)=∑h∈ℋi(t)v,h(t)d(v)= _h ^(t)_i N^(t)_v,h. Similarly, the degree of an edge h∈ℋi(t)h ^(t)_i is written as δ(h)=∑v∈ℐi(t)v,h(t)δ(h)= _v ^(t)_i N^(t)_v,h. Besides, we use i(t)∈ℕ|ℐi(t)|×|ℐi(t)| V^(t)_i ^|I^(t)_i|×|I^(t)_i| and i(t)∈ℕ|ℋi(t)|×|ℋi(t)| E^(t)_i ^|H^(t)_i|×|H^(t)_i| to be the diagonal matrices of the vertex degrees and edge degrees, respectively. Entity-based Hypergraph. To address the sparsity and limitations of historical user-item interaction data, we utilize the extensive DBpedia KG Auer et al. (2007) to construct the entity-based hypergraph. Specifically, we extract individual items mentioned in conversations as entities and their k-hop neighbors to form each hyperedge. This approach allows us to capture shared semantic connotations among the extended neighbors. The hyperedges are then connected based on the entities they have in common. Formally, the entity-based hypergraph entity(t)G^(t)_ entity can be represented as: entity(t) ^(t)_ entity =(ℐi(t)∪ℐe(t),ℋe(t),e(t)). =(I^(t)_i ^(t)_e,H^(t)_e, N^(t)_e). (3) where ℐe(t)I^(t)_e represents the k-hop neighbors, ℋe(t)H^(t)_e is the hyperedge set, and e(t)∈0,1|ℐe(t)|×|ℋe(t)| N^(t)_e∈\0,1\^|I^(t)_e|×|H^(t)_e| is the incidence matrix defined by Eq.(2). Similarly, e(t)∈ℕ|ℐe(t)|×|ℐe(t)| V^(t)_e ^|I^(t)_e|×|I^(t)_e| and e(t)∈ℕ|ℋe(t)|×|ℋe(t)| E^(t)_e ^|H^(t)_e|×|H^(t)_e| denote the diagonal matrices of vertex degrees and edge degrees, respectively. Word-based Hypergraph. Keywords in conversations are vital for understanding users’ needs. By analyzing prominent words, we can identify specific preferences, which is crucial for modeling diverse user preferences. To achieve this, we build a word-based hypergraph using the word-oriented KG ConcetNet Speer et al. (2017) to uncover semantic relations like synonymy, antonyms, and co-occurrence. We represent each historical conversation item as a keyword and extend it to include k-hop neighbors, forming a hyperedge. All the hyperedges connect through shared words. The word-based hypergraph word(t)G^(t)_ word can be defined as follows: word(t) ^(t)_ word =(ℐi(t)∪ℐw(t),ℋw(t),w(t)). =(I^(t)_i ^(t)_w,H^(t)_w, N^(t)_w). (4) where ℐw(t)I^(t)_w is k-hop neighbors, ℋw(t)H^(t)_w means the hyperedge set, and w(t)∈0,1|ℐw(t)|×|ℋw(t)| N^(t)_w∈\0,1\^|I^(t)_w|×|H^(t)_w| is the incidence matrix defined as Eq.(2). Similarly, let w(t)∈ℕ|ℐw(t)|×|ℐw(t)| V^(t)_w ^|I^(t)_w|×|I^(t)_w| and w(t)∈ℕ|ℋw(t)|×|ℋw(t)| E^(t)_w ^|H^(t)_w|×|H^(t)_w| denote the diagonal matrices of the vertex degrees and the edge degrees, respectively. 3.2.2 Multi-Preference Learning Upon constructing multiple hypergraphs as described earlier, we will leverage these hypergraphs to effectively capture diverse user preferences for mitigating the Matthew effect. This includes preferences related to items, entities, words, reviews, and knowledge aspects, all of which play a role in modeling multi-aspect preferences. Item-aspect Preference. Modeling item-aspect preferences holds significant importance in comprehending users’ distinct tastes and preferences concerning the various items they interact with. In line with this objective, we derive the item-aspect preference i P_i by leveraging the item-based hypergraph. To effectively capture high-order relations, inspired by Bai et al. (2021), we define our Hypergraph Convolution function HConv(⋅) HConv(·) as follows: m(l+1) X^(l+1)_m =HConv((l),i(t),i(t),i(t),i(t)), = HConv ( X^(l), N^(t)_i, V^(t)_i, E^(t)_i, I^(t)_i ), (5) HConv(⋅) HConv(·) =(i(t))−1i(t)(i(t))−1(i(t))T(l)i(l), =( V^(t)_i)^-1 N^(t)_i( E^(t)_i)^-1( N^(t)_i)^T X^(l) W^(l)_i, (l+1) X^(l+1) =Pooling(m(l+1))m=1M. = Pooling ( X^(l+1)_m )_m=1^M. Here, l X^l and (l+1) X^(l+1) represent the input of the l-th and (l+1)(l+1)-th layers, respectively, and i(l) W^(l)_i denotes the trainable parameter. The notations i(t) N^(t)_i, i(t) V^(t)_i, and i(t) E^(t)_i have been discussed in Section 3.2.1. Specifically, i(t) I^(t)_i signifies the item representations of ℐi(t)I^(t)_i extracted from the encoded entity embeddings Shang et al. (2023). Additionally, m denotes the number of heads in the multi-head architecture Vaswani et al. (2017c). Finally, we apply an average pooling Pooling(⋅) Pooling(·) on the representation (L+1) X^(L+1) obtained from the last layer (i.e., (L+1)(L+1) layer) to learn item-aspect preference i P_i: i=(L+1)=Pooling(m(L+1))m=1M, P_i= X^(L+1)= Pooling ( X^(L+1)_m )_m=1^M, (6) Entity-aspect Preference. Capturing entity-aspect preferences is highly advantageous for unveiling complex relationship patterns underlying users’ behaviors. To achieve this, we employ the entity-based hypergraph as a mechanism to learn entity-aspect preferences. In a similar vein to the item-aspect preference, the entity-aspect preference e P_e can be represented as: j(l+1) X^(l+1)_j =HConv((l),e(t),e(t),e(t),i+e(t)), = HConv ( X^(l), N^(t)_e, V^(t)_e, E^(t)_e, I^(t)_i+e ), (7) e P_e =(L+1)=Pooling(j(L+1))j=1J. = X^(L+1)= Pooling ( X^(L+1)_j )_j=1^J. The specifics of e(t) N^(t)_e, e(t) V^(t)_e, and e(t) E^(t)_e can be found in Section 3.2.1. Besides, i+e(t) I^(t)_i+e is the entity representations of the entity set ℐi(t)∪ℐe(t)I^(t)_i ^(t)_e. Moreover, j denotes the number of heads in the multi-head architecture, and e(l) W^(l)_e is the trainable parameter. Word-aspect Preference. Keywords occurring in conversations directly reflect users’ specific or potential preferences. Derived from the word-based hypergraph, the word-aspect preference w P_w can be formulated as: f(l+1) X^(l+1)_f =HConv((l),w(t),w(t),w(t),i+w(t)), = HConv ( X^(l), N^(t)_w, V^(t)_w, E^(t)_w, I^(t)_i+w ), (8) w P_w =(L+1)=Pooling(f(L+1))f=1F. = X^(L+1)= Pooling ( X^(L+1)_f )_f=1^F. Here w(t) N^(t)_w, w(t) V^(t)_w, and w(t) E^(t)_w are explained in more detail in Section 3.2.1. Additionally, i+w(t) I^(t)_i+w represents the word representations obtained from the encoded entity embeddings. The variable f denotes the number of heads in the multi-head architecture, and f(l) W^(l)_f denotes the trainable parameter. Review-aspect Preference. Item reviews provide valuable insights into users’ experiences and reflections. Analyzing these reviews helps identify patterns, sentiment trends, and user attitudes, leading to a better understanding of user preferences. Taking inspiration from the merits of Transformer model, we utilize the Transformer framework to encode accessed reviews Lu et al. (2021). Specifically, given a review R, the output embeddings from the previous transformer layer, denoted as l(R)T^l(R), define the subsequent layer l+1(R)T^l+1(R) using the Multi-head Attention function MHA(⋅) MHA(·) as: l+1(R)=MHA(l(R),l(R),l(R)), ^l+1(R)= MHA(T^l(R),T^l(R),T^l(R)), (9) MHA(,,)=[head1l;⋯;headgl]l, MHA( K, Q, V)=[ head^l_1;·s; head^l_g] W^l, headgl=SA(l(R)k,l(R)q,l(R)v), head^l_g= SA(T^l(R) W_k,T^l(R) W_q,T^l(R) W_v), SA(,,)=Softmax(Td/g), SA( K, Q, V)= Softmax( Q K T d/g) V, where g is the number of heads, l W^l denotes the trainable parameters, and each head headgl head^l_g is computed using the Scaled Dot-Product Attention Vaswani et al. (2017b) SA(⋅) SA(·). K, Q and V indicate the key, query and value matrices, respectively. k W_k, q W_q, and v W_v are learnable parameters. For convenience, we consider the output embeddings of the final transformer layer as the review-aspect preferences r P_r: r=MHA(ℒ(ℛ),ℒ(ℛ),L(ℛ)). P_r= MHA(T^L(R),T^L(R), T^L(R)). (10) Here ℒL is the number of transformer layers. Knowledge-aspect Preference. The information conveyed in the ongoing conversation reflects the dynamic preferences of the users, providing valuable insights into their current interests. Thus, our focus lies in modeling the preference for knowledge aspects by encoding the entities mentioned in the current conversation. Given the current conversation context C, we leverage DBpedia and CN-DBpedia, to extract entities ℰk=e1,e2,⋯,ekE_k=\e_1,e_2,·s,e_k\ along the paths. To capture high-order entity representations, we use RGCN to explicitly capture relational semantics by adopt contrastive pre-training Shang et al. (2023). The representation of entity e at the (l+1)(l+1)-th layer can be computed as: l+1=σ(∑r∈ℛ∑e^∈er1Zl1l^l+2ll), e^l+1=σ( _r _ e ^r_e 1Z_l W^l_1 e^l+ W^l_2 e^l), (11) where l e^l is the l-th layer’s representation of entity e, σ means the sigmoid function, e e refers to entities from the one-hop neighbor set erN^r_e under relation r, and ZlZ_l is the hyperparameter. 1l W^l_1 and 2l W^l_2 can be trained. We use the representation L e^L from the last layer as knowledge-aspect preference c P_c: c=RGCN(ℰk)=1T,2T,⋯,kT, P_c= RGCN(E_k)=\ e^T_1, e^T_2,·s, e^T_k\, (12) where iT e^T_i is the embedding of eie_i via RGCN. 3.3 Hypergraph-Aware CRS To combat the Matthew effect in the CRS, we adopt multi-aspect preferences, i.e., i P_i, e P_e, w P_w, r P_r, and c P_c, to accurately predict item in the recommendation task and effectively generate responses in the conversational task. 3.3.1 Recommendation Task The recommendation task aims to accurately predict items for users through natural conversations in dynamic user-system interactions. To address the Matthew effect, we first integrate multiple preferences to induce the fused preference mulrec P_ mulrec in the recommendation task as: h P_h =[i;e;w;r], =[ P_i; P_e; P_w; P_r], (13) mulrec P_ mulrec =Pooling([Pooling(h);c]). = Pooling([ Pooling( P_h); P_c]). where ; denotes the concatenation operation. Next, the vector mulrec P_ mulrec is used to select the suitable items in all the candidate set from item set ℐI, and the recommendation prediction is calculated as: rec=Softmax(mulrec⋅EIT), _ rec= Softmax( P_ mulrec· E^T_I), (14) where EI E_I is embeddings of all candidate items from item set ℐI. We use cross-entropy loss Shang et al. (2023) to learn the recommendation task: ℒr=−∑j=1B∑i=1|ℐ| _ r=-Σ^B_j=1Σ^|I|_i=1 [−(1−yij)⋅log(1−rec(j)(i)) [-(1-y_ij)· log(1-P^(j)_ rec(i)) (15) +yij⋅log(rec(j)(i))], +y_ij· log(P^(j)_ rec(i))], here the symbol B represents the size of the mini-batch, and yij∈0,1y_ij∈\0,1\ denotes the target label. 3.3.2 Conversational Task The conversation task focuses on generating proper dialogue utterances to respond to user inputs. To generate diverse responses, we integrate multi-aspect preferences vectors to derive the fused preference in the conversation task mulcon P_ mulcon as: mulcon P_ mulcon =MHA([c;h;h]), = MHA([ P_c; P_h; P_h]), (16) here h P_h is defined as Eq.(13). Then, this fused preference mulcon P_ mulcon is fed into the Transformer-based encoder-decoder framework for generating diverse responses. Let n−1 Y^n-1 be the output of the last time unit, then the current one n Y^n is: 0n A^n_0 =MHA(n−1,n−1,n−1), = MHA( Y^n-1, Y^n-1, Y^n-1), (17) 1n A^n_1 =MHA(0n,mulcon,mulcon), = MHA( A^n_0, P_ mulcon, P_ mulcon), 2n A^n_2 =MHA(1n,c,c), = MHA( A^n_1, P_c, P_c), 3n A^n_3 =MHA(1n,h,h), = MHA( A^n_1, P_h, P_h), 4n A^n_4 =β⋅2n+(1−β)⋅3n, =β· A^n_2+(1-β)· A^n_3, n Y^n =FFN(4n). = FFN( A^n_4). Here FFN(⋅) FFN(·) is the fully-connected feed-forward network, and β is hyper-parameter to balance two signals. To enhance the response diversity, we use preference-aware bias and item-related bias following Shang et al. (2023). Given the predicted sequence st−1\s_t-1\, the probability of the next token is calculated as: conv(st|st−1) _ conv(s_t|\s_t-1\) =P1(st|Yi)+P2(st|mulrec) =P_1(s_t|Y_i)+P_2(s_t| P_ mulrec) (18) +P3(st|mulrec), +P_3(s_t| P_ mulrec), where sts_t is the t-th utterances, and st−1=s1,s2,⋯,st−1\s_t-1\=s_1,s_2,·s,s_t-1. Inspired by Shang et al. (2023), P1(⋅)P_1(·), P2(⋅)P_2(·), and P3(⋅)P_3(·) are the vocabulary probability, vocabulary bias, and copy probability, respectively. Next, we use the cross-entropy loss: ℒcå _ cå =−∑b=1B∑t=1Tlog(conv(st|st−1)). =- _b=1^B _t=1^T log(P_ conv(s_t|\s_t-1\)). (19) Here T denotes the truncated length of utterances. Model REDIAL TG-REDIAL R@10 R@50 M@10 M@50 N@10 N@50 R@10 R@50 M@10 M@50 N@10 N@50 TextCNN 0.0644 0.1821 0.0235 0.0285 0.0328 0.0580 0.0097 0.0208 0.0040 0.0045 0.0053 0.0077 SASRec 0.1117 0.2329 0.0540 0.0593 0.0674 0.0936 0.0043 0.0178 0.0011 0.0017 0.0019 0.0047 BERT4Rec 0.1285 0.3032 0.0475 0.0555 0.0663 0.1045 0.0043 0.0226 0.0013 0.0020 0.0020 0.0058 ReDial 0.1705 0.3077 0.0677 0.0738 0.0925 0.1222 0.0038 0.0165 0.0012 0.0017 0.0018 0.0045 TG-ReDial 0.1679 0.3327 0.0694 0.0771 0.0924 0.1286 0.0110 0.0174 0.0048 0.0050 0.0062 0.0076 KBRD 0.1796 0.3421 0.0722 0.0800 0.0972 0.1333 0.0201 0.0501 0.0077 0.0090 0.0106 0.0171 KGSF 0.1785 0.3690 0.0705 0.0796 0.0956 0.1379 0.0215 0.0643 0.0069 0.0087 0.0103 0.0194 KGConvRec 0.1819 0.3587 0.0711 0.0794 0.0969 0.1358 0.0220 0.0524 0.0088 0.0102 0.0119 0.0185 BERT 0.1608 0.3525 0.0597 0.0688 0.0831 0.1255 0.0040 0.0194 0.0011 0.0017 0.0018 0.0050 XLNet 0.1569 0.3590 0.0583 0.0677 0.0811 0.1255 0.0040 0.0187 0.0011 0.0017 0.0017 0.0048 BART 0.1693 0.3783 0.0646 0.0744 0.0888 0.1350 0.0047 0.0187 0.0012 0.0017 0.0020 0.0048 MHIM 0.1966 0.3832 0.0742 0.0830 0.1027 0.1440 0.0300 0.0783 0.0108 0.0129 0.0152 0.0256 HyCoRec* 0.2231 0.4351 0.0797 0.0898 0.1123 0.1579 0.0377 0.0826 0.0154 0.0173 0.0162 0.0245 Table 1: Recommendation results. * indicates statistically significant improvement (p < 0.05) over all baselines. 4 Experiments and Analyses We conduct experiments to fully evaluate our HyCoRec and answer the following questions: • RQ1: How does HyCoRec perform compared with all baselines in the recommendation task? • RQ2: How does HyCoRec perform compared with all baselines in the conversation task? • RQ3: How does HyCoRec alleviate Matthew effect in the CRS? • RQ4: How do the item-based hypergraph item(t)G^(t)_ item, entity-based hypergraph enti(t)G^(t)_ enti, word-based hypergraph word(t)G^(t)_ word, and item reviews R contribute to the performance? • RQ5: How do parameters affect our HyCoRec? • RQ6: It is better to provide the case studies to comprehensively understand about how HyCoRec handles Matthew effect in the CRS? 4.1 Experimental Protocol Datasets. We evaluate our HyCoRec on two challenging CRS-based datasets REDIAL Li et al. (2018b) and TG-REDIAL Zhou et al. (2020b). The REDIAL consists of 11,348 dialogues involving 956 users and 6,924 items, while the TG-REDIAL contains 10,000 dialogues with 1,482 users and 33,834 items. The reviews in REDIAL are sourced from the IMDb†https://w.dbpedia.org/, while the reviews in TG-REDIAL are collected from Douban†https://movie.douban.com/. Baselines. To fully evaluate our HyCoRec, we conduct a comprehensive evaluation by comparing our method with several state-of-the-art methods. The compared methods include TextCNN Kim (2014), SASRec Kang and McAuley (2018), BERT4Rec Sun et al. (2019), Transformer Vaswani et al. (2017c), ReDial Li et al. (2018a), KBRD Chen et al. (2019), KGSF Zhou et al. (2020a), KGConvRec Sarkar et al. (2020), BERT Devlin et al. (2019), XLNet Yang et al. (2019), BART Lewis et al. (2020), DialoGPT Zhang et al. (2020), GPT-3 Brown et al. (2020), C2-CRS Zhou et al. (2022), LOT-CRS Zhao et al. (2023), UniCRS Deng et al. (2021b), and MHIM Shang et al. (2023). 4.2 Recommendation Performance (RQ1) Following Shang et al. (2023), we adopt Recall@K (R@K), MRR@K (M@K), NDCG@K (N@K) (K=10, 50) to evaluate the recommendation task. Experimental results in Table 1 validate that our HyCoRec outperforms all the compared methods. The improvement of HyCoRec over these baselines can be attributed to three reasons: (1) Incorporating external knowledge sources like DBpedia and ConceptNet into the CRS proves beneficial in exploring users’ intricate behaviors, considering the sparse and limited nature of user-item interaction data. (2) Dialogues serve as a treasure trove of valuable information beyond the explicit user inputs. By considering the ongoing conversation between the user and the system, HyCoRec can capture the user’s current context and understand their immediate needs. (3) Modeling multi-aspect preferences, including item-, entity-, word-, review-, and knowledge-aspect preferences, to enhance recommendation diversity and alleviate the Matthew effect as users interact with the system over time. 4.3 Conversational Performance (RQ2) In the conversational task, we adopt Distinct n-gram (Dist-n) Shang et al. (2023) (n=2,3,4) to evaluate the diversity of generated responses. Table 2 summarizes the experimental results, it is observed that our HyCoRec is superior to all the compared baselines. We can observe that the performance rankings of the four baseline models remain consistent, with KBRD leading the way, followed by KGSF, Transformer, and finally ReDial. This can be attributed to the fact that KBRD leverages external knowledge sources to align the representations of items and words. Besides, KGSF enriches its decoder by incorporating cross-attention mechanisms along with embeddings from both entity- and word-level knowledge graphs (KGs). Nevertheless, Transformer and ReDial solely rely on token sequences, disregarding the user preferences that are concealed within the entities. Compared with these baselines, the improvement of HyCoRec can be attributed to the fact that: (1) HyCoRec considers multi-aspect user preferences to generate diverse responses that effectively align with the user’s multi-level preferences for alleviating Matthew effect. (2) To accurately forecast the next utterance, we integrate the fused preference obtained from various aspects into a Transformer-based encoder-decoder framework to generate high-quality responses to meet users’ dynamic needs and interests. Model REDIAL TG-REDIAL Dist-2 Dist-3 Dist-4 Dist-2 Dist-3 Dist-4 ReDial 0.0214 0.0659 0.1333 0.2178 0.5136 0.7960 Trans. 0.0538 0.1574 0.2696 0.2362 0.7063 1.1800 KBRD 0.0765 0.3344 0.6100 0.8013 1.7840 2.5977 KGSF 0.0572 0.2483 0.4349 0.3891 0.8868 1.3337 C2-CRS 0.2623 0.3891 0.6202 0.5235 1.9961 2.9236 UniCRS 0.2464 0.4273 0.5290 0.6252 2.2352 2.5194 LOT-CRS 0.3312 0.6155 0.9248 0.9287 2.4880 3.4972 DialoGPT 0.3542 0.6209 0.9482 1.1881 2.4269 3.9824 GPT-3 0.3604 0.6399 0.9511 1.2255 2.5713 4.0713 MHIM 0.3278 0.6204 0.9629 1.1100 2.3520 3.8200 HyCoRec* 0.3661 0.6434 0.9523 1.2590 2.6000 4.1210 Table 2: Conversation results. * indicates statistically significant improvement (p < 0.05) over all baselines. 4.4 Study on Matthew Effect (RQ3) As our objective is to alleviate the Matthew effect in the CRS, we extensively examine the recommendation outcomes and compare them with the strongest baselines to assess whether HyCoRec can effectively mitigate Matthew effect. To mitigate the Matthew effect, the crucial factor is to enhance the diversity of the recommendation results. Thus, we adopt two commonly-used metrics Coverage@k (C@k) and Isolation-Index (Iso-Index) to assess the extent of recommendation diversification by taking into account the distinctions among recommended items. The higher coverage value demonstrates its superior capability to encompass a larger portion of the recommendation space, encompassing items from various categories. A lower isolation-index value indicates a higher diversity in the recommended results. As shown in Table 3, it is evident that our HyCoRec consistently achieves the highest values of Coverage and the lowest isolation-index value across all datasets compared with the strongest baselines. For instance, on the REDIAL, our HyCoRec achieves substantial improvements of 102.72%, 75.90%, 144.35%, and 63.75% in terms of Cover@5 when compared to all the strong models, namely KBRD, KGSF, KGConvRec, and MHIM, respectively. The results show that HyCoRec effectively addresses isolation and ensures extensive coverage of recommended items by providing users with a broader choice range, validating the superiority in alleviating the Matthew effect as the user interacts with the system. Datasets REDIAL Models C@5 C@10 C@15 C@20 Iso-Index KBRD 0.0579 0.0810 0.0961 0.1072 0.1149 KGSF 0.0664 0.0831 0.1195 0.1366 0.1055 KGConvRec 0.0478 0.0735 0.1044 0.1235 0.1003 MHIM 0.1098 0.1492 0.1747 0.1977 0.0923 HyCoRec 0.1168 0.1579 0.1848 0.2071 0.0617 Datasets TG-REDIAL Models C@5 C@10 C@15 C@20 Iso-Index KBRD 0.0757 0.1204 0.1468 0.1584 0.1222 KGSF 0.0847 0.1324 0.1606 0.1858 0.1198 KGConvRec 0.0720 0.0904 0.1228 0.1515 0.1091 MHIM 0.1749 0.2493 0.2939 0.3423 0.1042 HyCoRec 0.1841 0.2743 0.3100 0.3608 0.0791 Table 3: Results on C@k and Iso-Index metrics. Model REDIAL TG-REDIAL R@10 R@50 R@10 R@50 HyCoRec 0.2231 0.4351 0.0377 0.0826 w/o item hypergraph 0.2061 0.4177 0.0347 0.0733 w/o entity hypergraph 0.2049 0.4206 0.0267 0.0696 w/o word hypergraph 0.2058 0.4194 0.0257 0.0661 w/o item reviews 0.2102 0.4203 0.0267 0.0771 Table 4: Ablation studies on the recommendation task. 4.5 Ablation Studies (RQ4) In this part, we conduct ablation experiments with different variants of HyCoRec to verify the contributions of each component, including: 1) w/o item hypergraph: we remove the item-based hypergraph; 2) w/o entity hypergraph: we remove the entity-based hypergraph; 3) w/o word hypergraph: we remove the word-based hypergraph; 4) w/o item reviews: we remove item reviews. As shown in Table 4, we can observe that a substantial decline in performance when removing any type of component. The main reason is that various knowledge data can effectively explore users’ multi-aspect preferences. This observation highlights the effectiveness of HyCoRec in alleviating the Matthew effect by providing diverse recommendation results. 4.6 Hyperparameters Analysis (RQ5) Next, we investigate the impact of several important hyperparameters on the recommendation performance. As depicted in Fig.2, we can observe: Firstly, with the increase of embedding dimension, the recommendation performance continually improves. This is because the large dimension could encode sufficient high-level feature representations. Secondly, the model performance is optimal when the layer number is set to 2 on both datasets. The main reason is that larger hypergraph convolution layer numbers easily lead to model overfitting while the smaller one fail to capture enough feature representations. Lastly, a suitable hypergraph pooling layer number can enhance the model performance but the larger one might damage recommendation performance. The reason is that too large hypergraph pooling layer numbers might lose the important feature representations. Figure 2: Impact of different hyperparameters. 4.7 Case Studies (RQ6) For a more in-depth understanding of how our proposed method, HyCoRec, tackles the Matthew effect during user-system interactions, we present comparative case studies between our approach and existing methods, visually illustrating the dialogue recommendation outcomes in human-computer interaction. As illustrated in Fig.3, our method effectively recommends a diverse range of movies from different categories (see (a)), setting itself apart from existing methods that typically recommend movies from the same category (see (b)). The results highlight that our method achieves higher recommendation diversity, while most existing methods demonstrate lower diversity in recommendations. Generally, an effective strategy to alleviate the Matthew effect involves enhancing recommendation diversification Anderson et al. (2020); Hansen et al. (2021); Liang et al. (2021); Zheng et al. (2021b). Thus, these results validate the effectiveness of our proposed method in mitigating the Matthew effect as users interact with the system over time in CRS. Figure 3: Case studies to comprehensively understand about how our proposed method HyCoRec handles Matthew effect in the CRS compared with most existing methods. Different colors denote different categories (see (a)) while the same color means the same category (see (b)). 5 Conclusion The Matthew effect is a notorious issue in the CRS, and it will be increasingly amplified due to the dynamic user-system feedback loop. To address these issues, we propose a novel paradigm, HyCoRec, which aims to learn multi-aspect user preferences, i.e., item-, entity-, word-, review-, and knowledge-aspect preferences, to effectively generate diverse responses in the conversation task and accurately predict items in the recommendation task for alleviating Matthew effect. Extensive experiments validate that our HyCoRec outperforms all the compared baselines and the superior of HyCoRec in alleviating Matthew effect in the CRS. 6 Limitations While our HyCoRec has attained a remarkable state-of-the-art performance, it does have certain limitations. Firstly, the complexity and extensive nature of item reviews make the construction of the review-based hypergraph challenging and difficult. Consequently, the current version does not include the review-based hypergraph to capture a wider range of multiplex user relation patterns. Secondly, our proposed method necessitates the design of individual hypergraphs for learning multi-aspect preferences. This limitation could be addressed by developing a general framework that integrate any types of hypergraphs, thereby automatically unifying various knowledge sources. 7 Ethics Statement The data utilized in our study are sourced from open-access repositories, and do not pose any privacy concerns. We are confident that our research adheres to the ethical standards set forth by ACL. 8 Acknowledgements This work was supported in part by the National Key Research and Development Program of China under Grant No.2021ZD0111601; National Natural Science Foundation of China under Grant No.62325605, Grant No.62206110 and Grant No.62206314; Guangzhou Basic Research Project for Basic and Applied Research under Grant No.202201010334; Guangdong Basic and Applied Basic Research Foundation under Grant No.2023A1515011374 and Grant No.2022A1515011835; Guangzhou Science and Technology Program under Grant No.2024A04J6365; Science and Technology Projects in Guangzhou under Grant No.2024A04J4388; and Guangdong Province Key Laboratory of Information Security Technology, Sun Yat-sen University. References A. Anderson, L. Maystre, I. Anderson, R. Mehrotra, and M. Lalmas (2020) Algorithmic effects on the diversity of consumption on spotify. In The Web Conference, p. 2155–2165. Cited by: §1, §1, §2.2, §4.7. S. Auer, C. Bizer, G. Kobilarov, J. Lehmann, R. Cyganiak, and Z. G. Ives (2007) DBpedia: A nucleus for a web of open data. In International Semantic Web Conference/Asian Semantic Web Conference, Vol. 4825, p. 722–735. Cited by: §3.2.1. S. Bai, F. Zhang, and P. H. S. Torr (2021) Hypergraph convolution and hypergraph attention. Pattern Recognit. 110, p. 107637. Cited by: §3.2.2. T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei (2020) Language models are few-shot learners. In Conference on Neural Information Processing Systems, Cited by: §4.1. Q. Chen, J. Lin, Y. Zhang, M. Ding, Y. Cen, H. Yang, and J. Tang (2019) Towards knowledge-based recommender dialog system. arXiv preprint arXiv:1908.05391. Cited by: §2.1, §4.1. Y. Deng, Y. Li, F. Sun, B. Ding, and W. Lam (2021a) Unified conversational recommendation policy learning via graph-based reinforcement learning. In Conference on Research and Development in Information Retrieval, p. 1431–1441. Cited by: §2.1. Y. Deng, Y. Li, F. Sun, B. Ding, and W. Lam (2021b) Unified conversational recommendation policy learning via graph-based reinforcement learning. In Conference on Research and Development in Information Retrieval, p. 1431–1441. Cited by: §4.1. Y. Deng, W. Zhang, W. Xu, W. Lei, T. Chua, and W. Lam (2023) A unified multi-task learning framework for multi-goal conversational recommender systems. ACM Transactions on Information Systems 41 (3), p. 1–25. Cited by: §2.1. J. Devlin, M. Chang, K. Lee, and K. Toutanova (2019) BERT: pre-training of deep bidirectional transformers for language understanding. In the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, p. 4171–4186. Cited by: §4.1. E. V. Epure and R. Hennequin (2023) A human subject study of named entity recognition in conversational music recommendation queries. In European Chapter of the Association for Computational Linguistics, p. 1273–1288. Cited by: §1. C. Gao, K. Huang, J. Chen, Y. Zhang, B. Li, P. Jiang, S. Wang, Z. Zhang, and X. He (2023) Alleviating matthew effect of offline reinforcement learning in interactive recommendation. In Conference on Research and Development in Information Retrieval, p. 238–248. Cited by: §1. Y. Ge, S. Zhao, H. Zhou, C. Pei, F. Sun, W. Ou, and Y. Zhang (2020) Understanding echo chambers in e-commerce recommender systems. In Conference on Research and Development in Information Retrieval, p. 2261–2270. Cited by: §1. C. Hansen, R. Mehrotra, C. Hansen, B. Brost, L. Maystre, and M. Lalmas (2021) Shifting consumption towards diverse content on music streaming platforms. In Conference on Web Search and Data Mining, p. 238–246. Cited by: §1, §1, §2.2, §4.7. E. Hussein, P. Juneja, and T. Mitra (2020) Measuring misinformation in video search platforms: an audit study on youtube. ACM on Human-Computer Interaction 4 (CSCW), p. 048:1–048:27. Cited by: §3.2. X. Jin, W. Zhou, J. Wang, D. Xu, and Y. Zheng (2023) An order-complexity aesthetic assessment model for aesthetic-aware music recommendation. In Proceedings of the 31st ACM International Conference on Multimedia, p. 6938–6947. Cited by: §2.1. W. Kang and J. J. McAuley (2018) Self-attentive sequential recommendation. In IEEE International Conference on Data Mining, p. 197–206. Cited by: §4.1. Y. Kim (2014) Convolutional neural networks for sentence classification. In Empirical Methods in Natural Language Processing (Demonstrations), p. 1746–1751. Cited by: §4.1. W. Lei, X. He, Y. Miao, Q. Wu, R. Hong, M. Kan, and T. Chua (2020a) Estimation-action-reflection: towards deep interaction between conversational and recommender systems. In Web Search and Data Mining, p. 304–312. Cited by: §2.1. W. Lei, G. Zhang, X. He, Y. Miao, X. Wang, L. Chen, and T. Chua (2020b) Interactive path reasoning on graph for conversational recommendation. In International Conference on Knowledge Discovery and Data Mining, p. 2073–2083. Cited by: §2.1. M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V. Stoyanov, and L. Zettlemoyer (2020) BART: denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In the Association for Computational Linguistics, p. 7871–7880. Cited by: §4.1. R. Li, S. Ebrahimi Kahou, H. Schulz, V. Michalski, L. Charlin, and C. Pal (2018a) Towards deep conversational recommendations. Advances in Neural Information Processing Systems 31. Cited by: §4.1. R. Li, S. E. Kahou, H. Schulz, V. Michalski, L. Charlin, and C. Pal (2018b) Towards deep conversational recommendations. In Advances in Neural Information Processing Systems, p. 9748–9758. Cited by: §4.1. S. Li, Q. Cheng, L. Li, and X. Qiu (2023) Mitigating negative style transfer in hybrid dialogue system. In Association for the Advancement of Artificial Intelligence, p. 13103–13111. Cited by: §1. S. Li, R. Xie, Y. Zhu, X. Ao, F. Zhuang, and Q. He (2022) User-centric conversational recommendation with multi-aspect user modeling. In Conference on Research and Development in Information Retrieval, p. 223–233. Cited by: §2.1. Y. Liang, T. Qian, Q. Li, and H. Yin (2021) Enhancing domain-level and user-level adaptivity in diversified recommendation. In Conference on Research and Development in Information Retrieval, p. 747–756. Cited by: §1, §1, §2.2, §4.7. P. Liu, K. Shivaram, A. Culotta, M. A. Shapiro, and M. Bilgic (2021) The interaction between political typology and filter bubbles in news recommendation algorithms. In The Web Conference, p. 3791–3801. Cited by: §3.2. Y. C. Liu and M. Q. Huang (2021) Examining the matthew effect on youtube recommendation system. In Conference on Technologies and Applications of Artificial Intelligence, p. 146–148. Cited by: §1, §1, §2.2. Y. Liu, W. Zhang, B. Dong, Y. Fan, H. Wang, F. Feng, Y. Chen, Z. Zhuang, H. Cui, Y. Li, and W. Che (2023) U-NEED: A fine-grained dataset for user needs-centric e-commerce conversational recommendation. In Conference on Research and Development in Information Retrieval, p. 2723–2732. Cited by: §1. Y. Lu, J. Bao, Y. Song, Z. Ma, S. Cui, Y. Wu, and X. He (2021) RevCore: review-augmented conversational recommendation. In Findings of the Association for Computational Linguistics, p. 1161–1173. Cited by: §3.2.2. K. Mishra, P. Priya, and A. Ekbal (2023) Help me heal: A reinforced polite and empathetic mental health and legal counseling dialogue system for crime victims. In Association for the Advancement of Artificial Intelligence, p. 14408–14416. Cited by: §1. T. T. Nguyen, P. Hui, F. M. Harper, L. G. Terveen, and J. A. Konstan (2014) Exploring the filter bubble: the effect of using recommender systems on content diversity. In The Web Conference, p. 677–686. Cited by: §3.2. M. Qian, Y. Zheng, J. Qin, and L. Lin (2023) HutCRS: hierarchical user-interest tracking for conversational recommender system. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Singapore, p. 10281–10290. External Links: Link, Document Cited by: §2.1. L. Qin, Z. Li, Q. Yu, L. Wang, and W. Che (2023) Towards complex scenarios: building end-to-end task-oriented dialogue system across multiple knowledge bases. In Association for the Advancement of Artificial Intelligence, p. 13483–13491. Cited by: §1. X. Ren, H. Yin, T. Chen, H. Wang, Z. Huang, and K. Zheng (2021) Learning to ask appropriate questions in conversational recommendation. In Conference on Research and Development in Information Retrieval, p. 808–817. Cited by: §2.1. R. Sarkar, K. Goswami, M. Arcan, and J. P. McCrae (2020) Suggest me a movie for tonight: leveraging knowledge graphs for conversational recommendation. In Conference on Computational Linguistics, p. 4179–4189. Cited by: §4.1. C. Shang, Y. Hou, W. X. Zhao, Y. Li, and J. Zhang (2023) Multi-grained hypergraph interest modeling for conversational recommendation. AI Open 4, p. 154–164. Cited by: §2.1, §3.2.2, §3.2.2, §3.3.1, §3.3.2, §3.3.2, §4.1, §4.2, §4.3. R. Speer, J. Chin, and C. Havasi (2017) ConceptNet 5.5: an open multilingual graph of general knowledge. In Association for the Advancement of Artificial Intelligence, p. 4444–4451. Cited by: §3.2.1. H. Steck (2018) Calibrated recommendations. In Conference on Recommender Systems, p. 154–162. Cited by: §1. F. Sun, J. Liu, J. Wu, C. Pei, X. Lin, W. Ou, and P. Jiang (2019) BERT4Rec: sequential recommendation with bidirectional encoder representations from transformer. In International Conference on Information and Knowledge Management, p. 1441–1450. Cited by: §4.1. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. u. Kaiser, and I. Polosukhin (2017a) Attention is all you need. Advances in neural information processing systems 30. Cited by: §2.1. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin (2017b) Attention is all you need. In Advances in Neural Information Processing Systems, p. 5998–6008. Cited by: §3.2.2. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin (2017c) Attention is all you need. In Advances in Neural Information Processing Systems, p. 5998–6008. Cited by: §3.2.2, §4.1. H. Wang, Z. Wang, and W. Zhang (2019) Quantitative analysis of matthew effect and sparsity problem of recommender systems. CoRR. Cited by: §2.2. K. Xu, J. Yang, J. Xu, S. Gao, J. Guo, and J. Wen (2021) Adapting user preference to online feedback in conversational recommendation. In Web Search and Data Mining, p. 364–372. Cited by: §2.1. Z. Yang, Z. Dai, Y. Yang, J. G. Carbonell, R. Salakhutdinov, and Q. V. Le (2019) XLNet: generalized autoregressive pretraining for language understanding. In Advances in Neural Information Processing Systems, p. 5754–5764. Cited by: §4.1. Y. Zhang, F. Feng, X. He, T. Wei, C. Song, G. Ling, and Y. Zhang (2021) Causal intervention for leveraging popularity bias in recommendation. In Conference on Research and Development in Information Retrieval, p. 11–20. Cited by: §1. Y. Zhang, S. Sun, M. Galley, Y. Chen, C. Brockett, X. Gao, J. Gao, J. Liu, and B. Dolan (2020) DIALOGPT : large-scale generative pre-training for conversational response generation. In the Association for Computational Linguistics, p. 270–278. Cited by: §4.1. Z. Zhao, K. Zhou, X. Wang, W. X. Zhao, F. Pan, Z. Cao, and J. Wen (2023) Alleviating the long-tail problem in conversational recommender systems. In ACM Conference on Recommender Systems, p. 374–385. Cited by: §4.1. Y. Zheng, Z. Chen, J. Qin, and L. Lin (2024a) FacetCRS: multi-faceted preference learning for pricking filter bubbles in conversational recommender system. In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence, AAAI’24/IAAI’24/EAAI’24. External Links: ISBN 978-1-57735-887-9, Link, Document Cited by: §2.1. Y. Zheng, M. Qian, G. Wang, Y. Liu, Z. Chen, M. Mao, L. Lin, and K. Lam (2025a) HyperCRS: hypergraph-aware multi-grained preference learning to burst filter bubbles in conversational recommendation system. In Findings of the Association for Computational Linguistics: ACL 2025, Vienna, Austria, p. 5597–5608. External Links: Link, Document, ISBN 979-8-89176-256-5 Cited by: §2.1. Y. Zheng, J. Qin, P. Wei, Z. Chen, and L. Lin (2024b) CIPL: counterfactual interactive policy learning to eliminate popularity bias for online recommendation. IEEE Transactions on Neural Networks and Learning Systems 35 (12), p. 17123–17136. External Links: Document Cited by: §2.1. Y. Zheng, G. Wang, Y. Liu, and L. Lin (2024c) Diversity matters: user-centric multi-interest learning for conversational movie recommendation. In Proceedings of the 32nd ACM International Conference on Multimedia, M ’24, New York, NY, USA, p. 9515–9524. External Links: ISBN 9798400706868, Link, Document Cited by: §2.1. Y. Zheng, G. Wang, J. Qin, Z. Chen, J. Lin, P. Wei, L. Lin, and K. Lam (2026a) CIREC: causal intervention-inspired policy learning to mitigate exposure bias for interactive recommendation. IEEE Transactions on Knowledge and Data Engineering 38 (1), p. 123–137. External Links: Document Cited by: §2.1. Y. Zheng, P. Wei, Z. Chen, Y. Cao, and L. Lin (2021a) Graph-convolved factorization machines for personalized recommendation. IEEE Transactions on Knowledge and Data Engineering 35 (2), p. 1567–1580. Cited by: §2.1. Y. Zheng, P. Wei, Z. Chen, C. Tang, and L. Lin (2024d) Routing user-interest markov tree for scalable personalized knowledge-aware recommendation. IEEE Transactions on Neural Networks and Learning Systems 35 (10), p. 14233–14246. External Links: Document Cited by: §2.1. Y. Zheng, Z. Xie, G. Wang, Z. Liu, L. Lin, and K. Lam (2025b) Why multi-interest fairness matters: hypergraph contrastive multi-interest learning for fair conversational recommender system. In Findings of the Association for Computational Linguistics: ACL 2025, W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 25673–25684. External Links: Link, Document, ISBN 979-8-89176-256-5 Cited by: §2.1. Y. Zheng, R. Xu, Z. Chen, G. Wang, M. Qian, J. Qin, and L. Lin (2024e) HyCoRec: hypergraph-enhanced multi-preference learning for alleviating matthew effect in conversational recommendation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Bangkok, Thailand, p. 2526–2537. External Links: Link, Document Cited by: §2.1. Y. Zheng, R. Xu, Y. Ma, G. Wang, L. Lin, and K. Lam (2026b) Diversity is all you need: self-supervised hypergraph learning for mitigating popularity bias in conversational recommender system. In ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 16602–16606. Cited by: §2.1. Y. Zheng, R. Xu, G. Wang, L. Lin, and K. Lam (2024f) Mitigating matthew effect: multi-hypergraph boosted multi-interest self-supervised learning for conversational recommendation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Miami, Florida, USA, p. 1455–1466. External Links: Link, Document Cited by: §2.1. Y. Zheng, C. Gao, L. Chen, D. Jin, and Y. Li (2021b) DGCN: diversified recommendation with graph convolutional networks. In The Web Conference, p. 401–412. Cited by: §1, §1, §2.2, §4.7. Y. Zheng, C. Gao, X. Li, X. He, Y. Li, and D. Jin (2021c) Disentangling user interest and conformity for recommendation with causal embedding. In The Web Conference, p. 2980–2991. Cited by: §1, §2.2. K. Zhou, W. X. Zhao, S. Bian, Y. Zhou, J. Wen, and J. Yu (2020a) Improving conversational recommender systems via knowledge graph based semantic fusion. In International Conference on Knowledge Discovery and Data Mining, p. 1006–1014. Cited by: §2.1, §4.1. K. Zhou, Y. Zhou, W. X. Zhao, X. Wang, and J. Wen (2020b) Towards topic-guided conversational recommender system. In International Conference on Computational Linguistics, p. 4128–4139. Cited by: §4.1. Y. Zhou, K. Zhou, W. X. Zhao, C. Wang, P. Jiang, and H. Hu (2022) C2-crs: coarse-to-fine contrastive learning for conversational recommender system. In Web Search and Data Mining, p. 1488–1496. Cited by: §2.1, §4.1.