A Heterogeneous Information Network based Cross Domain Insurance Recommendation System for Cold Start Users
Ye Bi, Liqiang Song, Mengqiu Yao, Zhenyu Wu, Jianming Wang, Jing Xiao
AA Heterogeneous Information Network based Cross DomainInsurance Recommendation System for Cold Start Users
Ye Bi, Liqiang Song, Mengqiu Yao, Zhenyu Wu, Jianming Wang, Jing Xiao [email protected],{songliqiang537,yaomengqiu621,wuzhenyu447,wangjianming888,xiaojing661}@pingan.com.cn
Ping An Technology Shenzhen Co., Ltd
ABSTRACT
Internet is changing the world, adapting to the trend of internetsales will bring revenue to traditional insurance companies. Onlineinsurance is still in its early stages of development, where cold startproblem (prospective customer) is one of the greatest challenges.In traditional e-commerce field, several cross-domain recommen-dation (CDR) methods have been studied to infer preferences ofcold start users based on their preferences in other domains. How-ever, these CDR methods couldn’t be applied to insurance domaindirectly due to the domain’s specific properties. In this paper, wepropose a novel framework called a Heterogeneous informationnetwork based Cross Domain Insurance Recommendation (HCDIR)system for cold start users. Specifically, we first try to learn moreeffective user and item latent features in both source and targetdomains. In source domain, we employ gated recurrent unit (GRU)to module users’ dynamic interests. In target domain, given thecomplexity of insurance products and the data sparsity problem,we construct an insurance heterogeneous information network(IHIN) based on data from PingAn Jinguanjia, the IHIN connectsusers, agents, insurance products and insurance product propertiestogether, giving us richer information. Then we employ three-level(relational, node, and semantic) attention aggregations to get userand insurance product representations. After obtaining latent fea-tures of overlapping users, a feature mapping between the two do-mains is learned by multi-layer perceptron (MLP). We apply HCDIRon Jinguanjia dataset, and show HCDIR significantly outperformsthe state-of-the-art solutions.
CCS CONCEPTS • Applied computing → Online insurance ; •
Information sys-tems → Data mining ; •
Networks → Data path algorithms.
KEYWORDS
Insurance Recommendation, Heterogeneous Information Network,Cross-domain Recommendation, Cold Start Problem
ACM Reference Format:
Ye Bi, Liqiang Song, Mengqiu Yao, Zhenyu Wu, Jianming Wang, Jing Xiao.2020. A Heterogeneous Information Network based Cross Domain Insur-ance Recommendation System for Cold Start Users. In
Proceedings of the
Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than ACMmust be honored. Abstracting with credit is permitted. To copy otherwise, or republish,to post on servers or to redistribute to lists, requires prior specific permission and/or afee. Request permissions from [email protected].
SIGIR ’20, July 25–30, 2020, Virtual Event, China © 2020 Association for Computing Machinery.ACM ISBN 978-1-4503-8016-4/20/07...$15.00 https://doi.org/10.1145/3397271.3401426 (a) Home Page (b) Nonfinanacial Domain (c) Insurance Domain
Figure 1: Online shopping on Jinguanjia. (a) is homepage. (b)is nonfinancial domain, providing daily necessities. (c) is in-surance domain, providing various insurance products.
ACM,Xi’an, China, 11 pages. https://doi.org/10.1145/3397271.3401426
Internet is changing the world, every segment of the economy isexperiencing dramatic change and is having to respond to shiftsin the value chain, enhanced consumer power, and altered compet-itive cycles. Internet insurance adapted to the trend of economicboom in internet age for two main sectors. For supply side, internetinsurance overcomes the limitations of live sales and geography,increasing the customer base. For demand side, internet sales aremore acceptable to young people, who are the main consumersof insurance products. Adapting to the trend of internet sales willbring revenue to traditional insurance companies.Internet insurance is still in its early stages of development,where cold start problem (prospective customer) is one of the great-est challenges. For example, PingAn Jinguanjia, one of the mostpopular comprehensive applications (App) in China, which boastsmore than 100 million registered users, has nearly 90% cold startusers in insurance domain (i.e. these registered users didn’t buy anyinsurance products). This situation resulted from many reasons.First, insurance policies are so complex that ordinary users are rel-atively lack of knowledge to understand them. Besides, insuranceproducts are typically bought to be used for a long time period (e.g.one year in car insurance). Attracting prospective customers playsa critical role in buildup of the competitive edge for traditionalinsurance company. Under this circumstances, our motivation forcreating an online insurance recommendation system stems fromproviding personalize recommendations for prospective customers, a r X i v : . [ c s . I R ] J u l IGIR ’20, July 25–30, 2020, Virtual Event, China Ye Bi, Liqiang Song, Mengqiu Yao, Zhenyu Wu, Jianming Wang, Jing Xiao and then building customer loyalty. To our knowledge, there are notmany models about recommendation systems in insurance domain,some includes [14, 20, 22]. However, these methods treat insurancedomain and traditional e-commerce equally, neglecting productcomplexity and data sparsity problem in insurance domain.In this paper, we focus on PingAn Jinguanjia, one of the mostpopular comprehensive applications in China. In addition to tradi-tional e-commerce products (defined as nonfinancial products inthis paper), e.g. electronics, household supplies, etc., it also providesfinancial products like insurance products, investment services. Be-sides, each registered customer would be assigned with an agent,who can help with enquiries, offer recommendations. As mentionedabove, even though Jinguanjia has a big user group, it does not havea greater share of sales in online insurance. In other words, most ofthe registered users did not buy any insurance products, thoughthey have relatively abundant activities in nonfinancial domain.As a result, we could not get enough information in insurance do-main only. Traditional recommendation systems, like collaboratefiltering (CF) [1], sequential-based models [2], could not performeffectively in insurance domain, since most of users only have lessthan 2 interactions in a year. To obtain enough information and getmore accurate recommendation, PingAn company tries to use sideinformation form Jinguanjia App (the interaction behaviors formnonfinancial domain), but to little avail.Cross-domain recommendation (CDR) [5, 12, 16, 17], which aimsto improve the recommendation performance by means of trans-ferring information from the source domain to the target domain,is one of the promising ways to solve data sparsity and cold startproblem. These methods assume that there exists overlap in in-formation between users and/or items across different domains,and train a mapping function from the source-domain into thetarget-domain. So the key factor for CDR method is to learn morecomprehensive and accurate user representations in two domain.However, the complexity of the insurance products and the severedata sparsity hinder us from learning user representations in insur-ance domain as accurate as possible. As a result, we could not applyCDR methods into insurance and nonfinancial domain directly.In order to help the users understand the complex insurancepolicies and get user representations as comprehensive and accu-rate as possible, an insurance heterogeneous information network(IHIN) is constructed according to the data from Jinguanjia App. InIHIN, we define four types of nodes corresponding to users, agents,insurance products and insurance product properties, and six typesof edges denoting various types of relations between them. Graphconvolutional networks (GCN) [7] and Graph attention networks(GAT) [26] as powerful deep representation learning method forgraph data, has shown superior performance on recommendation.However, these methods apply identical aggregation function onvarious types of edges, and the number of neighbors grows expo-nentially as the layers stacked up, which prohibit these methodsperforming efficiently on HIN. To deal with heterogeneous infor-mation, many state-of-the-art models emerge and has been provedto be efficient [24, 29, 30]. R-GCNs [24] are developed to deal withhighly multi-relational data. HAN [29] designs a two level (node-level and semantic-level) attentions to generate node embeddingby aggregating features from meta-path based neighbors. Inspired by these models, we propose a novel framework calleda Heterogeneous information network based Cross Domain Insur-ance Recommendation (HCDIR) system for cold start users. Specif-ically, we first try to learn more effective user and item latentfeatures in both source and target domains. In source domain, usersinteractions are rich, we can easily get the consume sequence ofusers, so we employ gated recurrent unit (GRU) [2] to module users’dynamic interests. In target domain, given the complexity of insur-ance products and the data sparsity problem, we construct an IHINbased on data from Jinguanjia App, the IHIN connects users, agents,insurance products and insurance product properties together, giv-ing us richer underlying information. Then we employ three-level(relational, node, and semantic) attention aggregations to get userand insurance product representations. After obtaining the latentfeatures of the overlapping users, a feature mapping between thetwo domains is learned by multi-layer perceptron (MLP).In summary, our contributions in this paper are as follows: • To the best of our knowledge, this is the first work to combinecross-domain mechanism and heterogeneous informationnetwork to give personalized recommendations for cold startusers in insurance domain. • For the complexity of insurance products, we construct aheterogeneous information network, which contains fourtypes of nodes and six types of relations. And we employthree level aggregations over IHIN to learn more effectiveuser and item representations in insurance domain. • We conduct experiments on real-world recommendationscenarios, and the results prove the efficacy of HCDIR overseveral state-of-the-art baselines.
Our dataset is collected from one of the largest e-commerce plat-form PingAn Jinguanjia. As shown in Figure 1, Jinguanjia providesnot only nonfinancial products (traditional e-commerce products),but also financial products like insurance products, investment ser-vices. Besides, each registered customer would be assigned withan agent, who can help with enquiries, offer recommendations. Inthis paper, we aim at providing recommendations to prospectivecustomers by CDR method in insurance domain, the users we useare overlapping users, who have interactions in both insurancedomain and nonfinancial domain. Our dataset is collected withinthe time period from June 1st 2018 to May 31th 2019, the statisticsof which is shown in Table 1.
Table 1: Statistics of Our dataset.
IS-domain (Target domain) NF-domain (Source domain)
User Nodes 117,613
Users 117,613
Item Nodes 42
Items 19,266
Agent Nodes 90,377
User-Item Interactions 1,995,168
Insurance Property Nodes 35
User-Iten Relations 344,206
User-Agent Relations 97,343
Item-Property Relations 275
Nonfinancial Domain.
The nonfinancial domain contains pur-sue logs of nonfinancial products (daily necessities) including clothes,
Heterogeneous Information Network based Cross Domain Insurance Recommendation System for Cold Start Users SIGIR ’20, July 25–30, 2020, Virtual Event, China skincare products, fruits, etc. Each item is associated with a descrip-tion, illustrating category, function, and so on. Besides, we alsohave the interaction order of each user.
Insurance Domain.
The insurance domain contains short-terminsurance (coverage time less than one year) including illness in-surances, accident insurances, medical insurances, education in-surances and other kinds of insurances. To better learn user repre-sentations, we construct an insurance heterogeneous informationnetwork, which contains four types of nodes (user (U), agent (A),insurance product (I), insurance property (P)), and six types of rela-tions among them (U ↔ I: purchase and be purchased by ; U ↔ A: beserved by and serve ; I ↔ P: possess and be possessed by ). Insurancepolicies are very complex, even the same product, if two customersare in different age groups, they may pay different price. To betterdescribe insurance products, we choose 35 insurance properties(e.g. price, age limit, coverage time, etc.) that customers care mostabout as nodes in IHIN. Is it necessary to design a recommendation system specifically forcold start users in online insurance domain? To answer this, westart by investigating the following questions.
Q1. Is online insurance the tendency?
First, we might won-der that if users really buy insurance products online, or they mayhave been used to buying insurance products in the traditionalway. To answer this question, we calculate the number of insuredimprovements on Ping An Jinguanjia App from 2015 to 2019, whichare showed in Figure 2. From the results, we can observe that on-line insurance experienced explosive growth from 2016 to 2018, thenumber of insured orders on Jinguanjia jumped by over 2 timesin 2018 compared with that in 2015. However, the growth in 2019entered a bottleneck period, so it is urgent for insurance companyto adjust their operation patterns to the internet trend.
Figure 2: The Number of Insured Improvements w.r.t. 2015on Jinguanjia AppQ2. Do users’ behaviours in nonfinancial domain have in-fluence on their behaviors in insurance domain?
In order toinvestigate the implicit relationships between users’ behaviors ininsurance domain and nonfinancial domain, we define a metriccalled group-buy-ratio. For a given group, the group-buy-ratio isdefined by the number of people who buy insurance products onJinguanjia for the first time divided by the total number of peoplein this group.group-buy-ratio = number of people first buying insurancetotal number of people in the group We first select two groups of customers in Jinguanjia, regular cus-tomer group (RCG) includes customers who have bought only somenonfinancial products on Jinguanjia before, new customer group(NCG) concludes new registered customers, i.e. customers who did’tbuy anything. Then we calculate group-buy-ratio for the two groupin six months and summarize them in Figure 3. Figure 3: Group-buy-ratio of RCG and NCG.
From Figure 3, we can see that group-buy-ratio of RCG is higherthan that of NCG. There may be two reasons. Firstly, regular cus-tomers in nonfinancial domain might be more willing to trust Jin-guanjia, since they have shopping experience in the App. Moreover,as shown in Figure 1(b), most goods Jinguanjia provides are healthproducts, customers who buy those products may concern moreabout themselves. So, users’ behaviors in nonfinancial domain mayhelp us make recommendations in insurance domain. We note herethat group-buy-ratio in each month is new purchases rate in thegroup. Since insurance products are typically bought to be used fora long time period (e.g. one year for car insurance), the customersmay not buy them again in a short time period, so group-buy-ratioof RCG decreases in our statistic period. Even though, group-buy-ratio of RCG is higher than that of NCG.
Q3. Are insurance policies really complex?
In traditional e-commerce domain, for example, in clothes, customers only needto see picture to decided weather they need the cloth or not. Ininsurance domain, understanding items may require a considerablecognitive overload. For example, there are 11 main terms and 30subsidiary terms in “PingAn critical illness insurance clause”. Themain terms include responsibilities of insurance company, exemp-tion of insured liability, rights and obligations for both policy holderand insurer, etc. The subsidiary terms includes some explanation ofmedical term and exception of the insurance. In a word, insurancepolicies are complex, and the complexity can be summarized asnumerous contents and complicated terminology.
Q4. Are agents affecting users’ buying behaviors in insur-ance domain?
To investigate whether agents are affecting users’buying decisions, we define ask-buy-ratio, which is the numberof customers who buy online insurance after consulting the agentdividing by the total number of customers who have consulted. Wefirst divide the agents according to their communication frequencywith customers, and list the ask-buy-ratio of the top 5%, top 10%,and top 15% communication frequency agents in Table 2. From Ta-ble 2, we can see that different agents have different ask-buy-ratio,ask-buy-ratio of the top 5% communication frequency agents ismore than four times that of the top 15% communication frequencyagents. This indicates that if a customer is assigned with an agent in
IGIR ’20, July 25–30, 2020, Virtual Event, China Ye Bi, Liqiang Song, Mengqiu Yao, Zhenyu Wu, Jianming Wang, Jing Xiao the top 5% communication frequency group, he /she may be morelikely to buy online insurance products.
Table 2: Ask-buy-ratio of different agents. communication frequency order T+1 T+2 T+3top % % % top % % % top % % % To sum up, we have following findings. • Online insurance is becoming more and more popular, thoughtit is in its growth bottleneck. A special recommendation sys-tem for online insurance domain is in demand. • Users’ behaviors in nonfinancial domain have influence ontheir behaviors in insurance domain. Users who have shop-ping experiences in nonfinancial domain are more willing totrust Jinguanjia, and more likely to buy insurance products. • Insurance policies are too complex to understand, the tradi-tional randomly initialized method could not give the accu-rate item representations. • Different agents have different influence on users, a userassigned with the top 5% agent will be more likely to buyonline insurance.Given the above findings, we argue that designing a recommen-dations system specifically for online insurance is essential. It isalso worth noting that, to give more accurate recommendations,we should try to represent the products accurately and take theinfluence of nonfinancial domain and agents into consideration.
A heterogeneous information network (HIN) is a special kind of in-formation network, which contains either multiple types of objectsor multiple types of relations, which can be defined as follows:
Definition 2.1 (Heterogeneous Information Network (HIN) [25]).
A HIN is defined as a directed graph G = (V , E) with an nodetype mapping function ϕ : V → A and a relation type mappingfunction φ : E → R . A and R denote the sets of predefined nodeand relation types, where |A| + |R| > Definition 2.2 (Meta-path [25]).
A meta-path ρ is defined as apath in the form of e r −−→ e r −−→ . . . r L − −−−→ e L (abbreviated as e , e , . . . , e l ), which describes a composite relation r = r ◦ r ◦ . . . ◦ r L − between object e and e L , where ◦ denotes the compositionoperator on relations. In this section, we formally define our problem, and summarizethe notations and descriptions in Table 3. As mentioned above, wehave two domains, a source domain (nonfinancial domain) anda target domain (insurance domain). Let U = { u , u , . . . , u m } denote overlapping users between nonfinancial domain D s andinsurance domain D t , respectively. If a user only appears in onedomain, he/she is a cold start user in the other domain. The user-item interaction matrices are denoted as Y s ∈ R m × s and Y t ∈ Table 3: Notations and descriptions
Notations Descriptions D s , D t source domain and target domain U overlapping users in the two domains Y s , Y t rating matrices of source and target domain NI su , NI tu interacted item sequences of user u in sourceand target domain N ( e ) e ’s one-hop neighbors N ρ ( e ) the set of nodes connecting to e by meta-path ρ R m × t , which are defined according to usersâĂŹ implicit feedbacks.We additionally use NI su and NI tu for the sequences of items thatuser u has interacted with. Besides, the interactions in insurancedomain can be abstracted as a heterogeneous information network(HIN), which we will illustrate later. Given rating matrices and HIN,our goal is to learn more effective latent features for users and items,and then learn the mapping function from nonfinancial domain toinsurance domain, which can help us deal with cold start users. To provide recommendations to cold start users, we propose HCDIR.As shown in Figure 4, HCDIR contains three main parts: learninglatent features of users in both insurance domain and nonfinancialdomain, mapping of user latent features.
Figure 4: The Framework of HCDIR
As mentioned above, the complexity of insurance products is typi-cally non-trivial, understanding the items may require a consider-able cognitive overload [22]. Under this circumstance, generatingefficient user embeddings is challenging. To achieve this goal, wedesign a three-level attention aggregation HIN method (TAHIN).In this part, we first introduce the IHIN we constructed based onJinguanjia dataset, and then present how to learn effective user rep-resentations over the constructed IHIN. Figure 5 shows the detailsof TAHIN module. Specifically, we first propose relational atten-tion to aggregate one-hop heterogeneous neighbors, and then nodeattention to aggregate meta-paths based neighbors, and semanticattention to aggregate meta-paths based neighbor sets. Finally, weaggregate the results of relational attention aggregation and seman-tic attention aggregation to the original node embedding to updatenode representations.
Interactions in insurance domain can be abstracted as an insuranceheterogeneous information network (IHIN). Specifically, we define
Heterogeneous Information Network based Cross Domain Insurance Recommendation System for Cold Start Users SIGIR ’20, July 25–30, 2020, Virtual Event, China
Figure 5: The Details of TAHIN (take node U as example). (a) illustrates the HIN constructed on Jinguanjia dataset. (b) isrelational neighbors aggregation, we first project the neighbors to the same node type space, and aggregate them by calculatingthe weighted sum of one-hop neighbors. (c) is the node and semantic attention aggregation, the left part is to aggregate themeta-paths based neighbors, the right part is to aggregate the results from the left part. (d) is node updation, aggregatinginformation from (b) and (c) to the original node representation. four types of nodes corresponding to user (U), agent (A), insuranceproduct (I) and insurance product property (P), and six types ofedges denoting various types of relations between them. As wementioned above, insurance products are very complex, customersusually couldn’t understand the whole insurance policies by justreading insurance products titles. Insurance products have severalproperties, which meet different demands for different customers.Therefore, we treat insurance property as a type of node. In this pa-per, we choose several properties customers most care about, whichare price, level of assurance, character, coverage time, insurancetype, age restriction, extra characters, etc. The schema of IHIN isdisplayed in Figure 5(a), which is formally defined as: Definition 4.1 (Insurance Heterogeneous Information Network).
In-surance Heterogeneous Information Network (IHIN) in our workis a HIN, containing four types of nodes: users e u , agents e a , insur-ance products e i and insurance product properties e p . Edges exitbetween e u and e a denoting be served by r ua and serve r au relations,between e u and e i denoting purchase r ui and be purchased by r iu relations, between e i and e p denoting possess r ip and be possessedby r pi relations.PingAn company possesses the data of user portrait, agent por-trait and item portrait, for efficiency, we initialize IHIN using thesedata instead of initializing them randomly.In IHIN, two nodes can be connected via different meta-paths.As shows in Figure 5, two insurance products can be connected viamultiple meta-paths, e.g. insurance product-user-insurance product(I-U-I), insurance product-insurance property-insurance product(I-P-I), etc. Different meta-paths may reveal different semantics. Forexample, I-U-I means the two insurance products are needed by the same user, they may be complementary. I-P-I means the twoinsurance products have same properties, e.g. high level assurance.In addition, meta-paths can also connect different types of nodes,for example, user-insurance product-insurance property (U-I-P),which implies that the user bought the insurance, since she mayconcern most about the insurance property. Now, we can give thedefinition of meta-path based neighbors: Definition 4.2 (Meta-path based Neighbors [10]).
Given a node e and a meta-path ρ in a HIN, the meta-path based neighbors N ρ ( e ) of node e is defined as a set of nodes which connect with node e via meta-path ρ . Note that the node’s meta-path based neighborsmay have different node types. As different relations implydifferent information, as shown in Figure 5(a), I and A are allneighbors of U , but they imply different information. We employa relational attention aggregation over one-hop neighbors. Figure5 (b) illustrates the framework. Specifically, instead of using thesame aggregation function among different one-hop neighbors, welearn a specific aggregation function for each type of relation. Let h e denote the current embedding of node e , as node’s one-hopneighbors may have different node type with the node, so we firstproject them to the same node space ( P r is projection matrix), andthen calculate the attention score: α ew = exp ( f r ( h e , P r h w )) (cid:205) j ∈N ( e ) exp ( f r ( h e , P r h j )) , ∀ w ∈ N ( e ) , where f r (· , ·) is the deep neutral network performing relationalattention, α ew is the level of influence of node e w , N ( e ) is node e ’s IGIR ’20, July 25–30, 2020, Virtual Event, China Ye Bi, Liqiang Song, Mengqiu Yao, Zhenyu Wu, Jianming Wang, Jing Xiao one-hop neighbors. Then, we aggregate information from N ( e ) : h e = σ (cid:169)(cid:173)(cid:171) (cid:213) w ∈N ( e ) α ew h w (cid:170)(cid:174)(cid:172) , (1)where σ denotes the activation function. Two nodes can also be con-nected by meta-paths, since different meta-path based neighborsimply different information (e.g. information from insurance prod-uct I to insurance product I ( I - U - I ) is different from informationfrom it to I ( I - P - I ), as shown in figure 5(a), since the formerimplies the same user, and the later implies the same property). Forefficiency, we only choose the meta-path based neighbors that havethe same node type with the node. The attention score is definedas: β ρew = exp ( f ρ ( h e , h w )) (cid:205) j ∈N ρ ( e ) exp ( f ρ ( h e , h j )) , ∀ w ∈ N ρ ( e ) , where f ρ (· , ·) is the deep neutral network which performs the node-level attention, β ew is the level of influence of node e w . Then, weemploy attention mechanism to aggregate the information of themeta-path based neighbors: h ρe = σ (cid:169)(cid:173)(cid:171) (cid:213) w ∈N ρ ( e ) β ρew h w (cid:170)(cid:174)(cid:172) , where σ denotes the activation function. The procedure is illus-trated in the left part of Figure 5(c).Given the meta-path set { ρ , ρ , . . . , ρ N } , after node attentionaggregation, for node e , we can obtain N node-level embeddings,denoted as { h ρ e , h ρ e , . . . , h ρ N e } . All the node embeddings are de-noted as { E ρ , E ρ , . . . , E ρ N } . In the following part, we introducehow to aggregate these node-level embeddings.To learn a more accurate node embedding, we try to fuse multiplenode embeddings. Taking { E ρ , E ρ , . . . , E ρ N } as input, as shownin the right part of Figure 5(c), we first calculate the importance ofeach meta-path ρ j : w ρ j = |V| (cid:213) e ∈V q T · tanh ( W h ρ j e + b ) , and the weight for ρ j , ( j = N ) is defined as: γ ρ j = exp ( w ρ j ) (cid:205) Nj = exp ( w ρ j ) . Form the definition of the attention score, we can see that the higher γ ρ j , the more important meta-path ρ j is. Now, we can fuse thesenode-level embeddings to obtain the final node embeddings: h e = N (cid:213) j = γ ρ j · h ρ j e . (2) Finally, we aggregate the information tonode e from h e (from (1)) and h e (from (2)): h e = ReLU { W concat [ h , W concat ( h e , h e ) + b ] + b } . After updating the HIN node embeddings, we can get the userand insurance product embedding, which are denoted as u t and v t ,respectively. The objective function in target domain is: L T = (cid:213) ( u , v )∈ Y t − ( y uv log ˆ y uv + ( − y uv ) log ( − ˆ y uv )) , (3)where ˆ y uv = σ ( f ( u t , v t )) , σ (·) is sigmoid function, f is a rankingfunction which can be a dot-product or a deep neural network. In Jinguanjia, each item i in nonfinancial domain is associated witha description c i . In order to learn more effective latent features, weemploy word2vec [18]. Suppose there are n words in i ’s content c i . Then we utilize word2vec to obtain word vectors, which arerepresented as { w ik } nk = . Then we concatenate word vectors andapply a max pooling over it to get the final item embedding: i = max-pooling ( concat ({ w ik } nk = )) . To model the final user latent feature u s , we employ GRU overthe user’s interacted sequence NI su , x n = σ ( W x i n + U x h n − + b x ) r n = σ ( W r i n + U r h n − + b r ) (cid:101) h n = tanh ( W h i n + r n ◦ U h h n − + b h ) h n = ( − x n ) ◦ h n − + x n ◦ (cid:101) h n , where σ is sigmoid function, ◦ is element-wise product, W x , W r , W h ∈ R n H × d , U x , U r , U h ∈ R n H × n H , n H = d is hidden size. Anduse h n to represent the user, i.e. u s = h n . The loss function is thesame as eq. (3), where ˆ y uv = σ ( f ( u s , v s )) . Similar to study [17], we employ MLP to perform latent spacematching from source domain to target domain. We take u s asinput and u t as output. and the loss function can be formalized as: L cross = (cid:213) u ∈U ∥ f mlp ( u s ) − u t ∥ . In this paper, we assume cold start users have interactions in non-financial domain, but no interactions in insurance domain. Afterlearning the latent features in nonfinancial domain u s , we can getthe corresponding mapping latent features ˆ u t = f mlp ( u s ) . Basedon learned ˆ u t , we can make recommendations to cold start users. To evaluate the performance of HCDIR, we conduct extensive ex-periments and online A/B test on Jinguanjia dataset to answer thefollowing key questions:
RQ1 : How does our proposed HCDIR model perform comparedwith the state-of-the-art methods for CDR task?
RQ2 : Can the proposed HCDIR alleviate the data sparsity prob-lem in the target domain?
RQ3 : How does different types of heterogeneous auxiliary in-formation and other HIN options affect the recommendation per-formance in HCDIR?
Heterogeneous Information Network based Cross Domain Insurance Recommendation System for Cold Start Users SIGIR ’20, July 25–30, 2020, Virtual Event, China
As described in Section 2.1, we build and release a suitable datasetfor insurance product recommendation task. We randomly split theoverlapping users of Jinguanjia dataset into training set (60%) tolearn parameters, validation set (20%) to tune hyper-parameters,and testing set (20%) for the final performance comparison. For thetesting set, we remove their information in the target domain toutilize them as cold start users for evaluating the recommendationperformance (i.e., test users). To study the performance changes ofour proposed methods with respect to the number of overlappingusers, we restrict the number of the overlapping users similarlyto the real-world distribution. We build four training sets with acertain fraction η ∈ { , , , } of the overlapping userswho do not belong to the test users in baseline comparison study. Four widely used recommendation algorithms are compared withthe variants of HCDIR. These baselines can be divided into twogroups: (1)
Single-domain Models : BPR [21] and GRU4REC [8]; (2)
Cross-domain Models : EMCDR-BPR [17], EMCDR-GRU, two vari-ants of HCDIR and HCDIR. The first group is utilized to validatethe usefulness of cross-domain recommendation models, and thesecond group is used to demonstrate the advantage of TAHIN mod-ule in insurance domain to deal with various kinds of heteroge-neous information including user purchase logs, agent and complexinsurance products’ properties. How to utilize different types ofheterogeneous information is one of the key factors to boost theeffectiveness of model. RGCN and HAN are two representativemethods in handling heterogeneous data. Here, we designed twovariants, HCDIR-RGCN and HCDIR-HAN adopting RGCN andHAN, respectively. HAN is superior to the other deep heteroge-neous network embedding models such as Metapath2Vec. RGCNemploys the relation-aggregators-based GCN to heterogeneousinformation network. HCDIR both leverages HAN and RGCN toprocess various kinds of heterogeneous information in insurancedomian for better user representation, which can effectively im-prove the recommendation performance.We evaluate all models with NDCG and Rec@N (N=1,3,5), whicheffectively evaluate the performance of recommendation methods.NDCG is used to observe the overall performance in terms of rank-ing insurances, while Recall@N is used to judge how accuratelyrecommend insurances at top N positions.
NDCG : Normalized Discounted Cumulative Gain (NDCG) ex-tends HR by assigning higher scores to the hits at higher positionsin the ranking list.
Recall@N(Rec@N) : The primary evaluation metric is Recall,which measures the proportion of cases when the relevant item isamongst the top ranked items in all test cases.
Parameter Setting.
In TAHINâĂŹs ‘relational neighbor aggrega-tionâĂŹ part, message passing is set as mean operation and typewise reducer is set as sum operation. In TAHINâĂŹs ‘meta-pathbased aggregationâĂŹ part, the meta-paths used here can be cate-gorized into four groups according to node types. User meta-pathsare [U I U], [U A U] and [U I P I U], item meta-paths are [I U I], [I P
Table 4: Performance comparison.
Jinguanjia dataset Metrics η Group Method NDCG Rec@1 Rec@3 Rec@510 % Single-domain BPR 0.0719 0.0213 0.0737 0.1248RS GRU4REC 0.0036 0.0017 0.0032 0.0057EMCDR-BPR 0.0881 0.0324 0.0689 0.1543Cross-domain EMCDR-GRU 0.1013 0.0284 0.0961 0.2088RS HCDIR-RGCN 0.2468 0.0967 0.3448 0.3849HCDIR-HAN 0.3206 0.1236 0.3476 0.4828HCDIR 0.3674 0.1366 0.4002 0.554320 % Single-domain BPR 0.0789 0.0241 0.0864 0.1348RS GRU4REC 0.0042 0.0022 0.0047 0.0061EMCDR-BPR 0.0984 0.0347 0.0848 0.1611Cross-domain EMCDR-GRU 0.1112 0.0366 0.1308 0.2257RS HCDIR-RGCN 0.2579 0.1003 0.3516 0.4002HCDIR-HAN 0.3311 0.1273 0.3656 0.4927HCDIR 0.3769 0.1548 0.4189 0.568350 % Single-domain BPR 0.0791 0.0274 0.1205 0.1735RS GRU4REC 0.0117 0.0027 0.0114 0.0213EMCDR-BPR 0.1125 0.0402 0.1609 0.2281Cross-domain EMCDR-GRU 0.1289 0.0496 0.1594 0.2589RS HCDIR-RGCN 0.2701 0.1166 0.3611 0.4341HCDIR-HAN 0.3432 0.1341 0.3946 0.5372HCDIR 0.3895 0.1636 0.4354 0.5827100 % Single-domain BPR 0.1009 0.0354 0.1627 0.1809RS GRU4REC 0.0137 0.0054 0.0154 0.0221EMCDR-BPR 0.1359 0.0511 0.2059 0.2556Cross-domain EMCDR-GRU 0.1498 0.0806 0.2124 0.2486RS HCDIR-RGCN 0.3067 0.1247 0.3739 0.4974HCDIR-HAN 0.3703 0.1357 0.4254 0.5627HCDIR 0.4109 0.1873 0.4654 0.6128
I] and [I U A U I], agent meta-paths are [A U A] and [A U I U A]and insurance productsâĂŹ properties meta-paths are [P I P] and[P I U I P]. The number of attention head in GAT is set to 8. Owingto separate training in three tasks (insurance domain, nonfinan-cial domain and cross domain) in cold start scenario, single typeof meta-paths cannot significantly affect the model performancewhile incorporation of all kinds of meta-paths can boost the perfor-mance. Final embedding dimension S is a key parameter in HCDIRdiscussed in the below section. We set the GRU hidden state size to32 due to storage. We take Adam as our optimizing algorithm. Forthe hyper-parameters of the Adam optimizer,we set the learningrate α = 0.001. These settings are chosen with grid search on thevalidation set. To speed up the training and converge quickly, weuse batch size as 32. We test the model performance on the vali-dation set for every epoch. We implement the proposed methodbased on Pytorch and DGL [27]. All experiments are performed inNvidia Tesla V100. Study of the Final Embedding Dimension S.
The quality ofthe final emvedding can directly effect the performance of model.As shown in Figure 6, we can see that with the increase of theembedding dimension S, the performance raises first and then startsto drop slowly. The best parameter of S is 32. The main reason is thatcross-domain method needs a suitable dimension to encode twodomains’ different information and larger dimension may introduceadditional redundancies.
To answer RQ1 and RQ2, two variants of HCDIR are compared withfour state-of-the-art models with different densities. Table 4 showsthe performance comparison. Overall, benefiting from the proposedTAHIN module and source domain information, HCDIR beats allcomparative methods under all levels of data sparsity, respectively.These experiments reveal a number of interesting discoveries: (1)
IGIR ’20, July 25–30, 2020, Virtual Event, China Ye Bi, Liqiang Song, Mengqiu Yao, Zhenyu Wu, Jianming Wang, Jing Xiao
Figure 6: Study of the Final Embedding Dimension S.
All cross-domain methods yield better performances than single-domain methods with mixture of target and source domain data ,demonstrating the importance of cross-domain module; (2) Owingto the capability of using different types of heterogeneous informa-tion in insurance domian, two variants of HCDIR (HCDIR-RGCNand HCDIR-HAN) defeat other comparative methods; (3) HCDIRachieves a better performance in a sparser dataset compared withother methods. It is validated that, compared to comparative ap-proaches, HCDIR can better alleviate the negative impacts of thedata sparsity issue.In order to anwser RQ3, we conduct experiments to compareHCDIR with HCDIR-RGCN and HCDIR-HAN. From the results ofTable 4, we can find that the performance of HCDIR-RGCN andHCDIR-HAN declines sharply in terms of all the metrics whenusing a sparser dataset. This experiment shows that, the proposedHCDIR can get more stable and better performance with limiteddata, which mainly contributes various types of heterogeneousinformation and the incorporation of RGCN and HAN to deal withvarious kinds of auxiliary relationships.
We find two important factors (data and corresponding data processmodule) effecting the performance of model. Therefore, we conductthe following two studies at 10% sparsity level, data ablation andmodel ablation study as shown in Table 5.
Result 1: Data Ablation
In order to investigate the effect of two newly added heteroge-neous data (Agents and Insurance Properties), we designed threevariants of our proposed model, HCDIR with only interactions,HCDIR without Agent and HCDIR without Insurance Properties(short for HCDIR without IP). From Table 5, it is found that onlyusing interactions can not reach the best performance even use ourproposed model framework. We can also observe that the perfor-mance of HCDIR without Agent declines more than HCDIR withoutIP compared to HCDIR in terms of all the metrics, which meansagent heterogeneous information is the key factor to improve themodel.
Result 2: Model Ablation
Two kinds of newly added heterogeneous information are usedin HCDIR. How to leverage various types of heterogeneous infor-mation effectively may affect the final model performance. RGCNand HAN are two widely used methods in dealing with heteroge-neous data, so we designed two variants of HCDIR in IHIN module,
Table 5: Performance of variants of HCDIR on Jinguanjiadataset at 10 % sparsity level inguanjia dataset Metricsat 10 % sparsity level NDCG Rec@1 Rec@3 Rec@5 HCDIR only 0.1013 0.0284 0.0961 0.2088with interactions -72.43 % -79.21 % -75.99 % -62.33 % Data
HCDIR 0.2157 0.0933 0.2287 0.3277Ablation without Agent -41.29 % -31.70 % -42.85 % -40.88 % HCDIR 0.2313 0.1073 0.2512 0.3665without IP -37.04 % -21.45 % -37.23 % -33.88 % HCDIR 0.2468 0.0967 0.3448 0.3849Model using RGCN -32.83 % -29.21 % -13.84 % -30.56 % Ablation
HCDIR 0.3206 0.1236 0.3476 0.4828using HAN -12.74 % -9.52 % -13.14 % -12.90 % Full Model
HCDIR 0.3674 0.1366 0.4002 0.5543 namely HCDIR using RGCN and HCDIR using HAN. HCDIR usingHAN outperforms HCDIR using RGCN which aggregates 1-hoprelation-aware neighbors. These results indicates the advantageof the combination attention mechanism and higher-order hetero-geneous neighbors generated by GCN-based model in HAN. Toimprove HCDIR, we choose HAN and RGCN to deal with theseheterogeneous information, which gains a better result.
To validate the effectiveness of HCDIR, we implement online A/Btest for insurance domainâĂŹs cold-start users to show how crossdomain method and heterogeneous insurance information affectcold start recommendation in real-world scenario.For online A/B testing, cold-start users who havenâĂŹt pur-chased any insurance products by the end of August 2019 are di-vided into three groups with highly similar activities in JinguangjiaAPP where each group contains 150,000 users. Users of the firstgroup users are recommended insurance products by traditionalstrategy using best trained machine learning model LightGBM, do-nated as G_Baseline. Users of the second group are recommendedby HCDIR without agent heterogeneous information with 10% train-ing data of Jinguanjia dataset used above, donated as G_HCDIRwithout agent. Users of the third group are by our proposed HCDIRtrained with 10% training data, donated as G_HCDIR.
User Purchase Conversion Rate (UPCR): Number of users whopurchased the recommended insurance product divide total numberof cold start users
User Premium Growth Amount (UPGA): Amount of insurancepremium cold start users paid for the recommended insurance.Table 6 shows the results of our designed online A/B testingcompared with G_Baseline as baseline. We compare the perfor-mance of G_HCDIR without agent to machine learning methodG_Baseline using LightGBM with only user-item interactions anddesigned features. From Table 6, we find that the performance ofG_HCDIR without agent and G_HCDIR using all the heterogeneousinformation consistently outperform these baseline methods. Theimprovement of UPCR and UPGR gradually increase over time,which indicates it need time for cold start users to develop insur-ance awareness. Specifically, it can be observed that G_HCDIRwithout agent at least improves UPCR and UPGA by 8.79 % and10.97 % in the time period from 1 month to 3 months compared
Heterogeneous Information Network based Cross Domain Insurance Recommendation System for Cold Start Users SIGIR ’20, July 25–30, 2020, Virtual Event, China
Table 6: Online performance of compared methods. âĂŸG_BaselineâĂŹ indicates the baseline performance of cold start usergroup using traditional method LightGBM; and âĂŸG_HCDIR without agentâĂŹ and âĂŸG_HCDIRâĂŹdenotes HCDIR with-out agent heterogeneous relationships and HCDIR, respectively.
Metrics G_HCDIR without agent G_HCDIRT+1 month T+2 months T+3 months T+1 month T+2 months T+3 monthsimprovement percentage 8.79% 12.87% 18.38% 12.94% 18.66% 23.20%of UPCR vs G_Baselineimprovement percentage 10.97% 13.04% 15.31% 15.25% 20.41% 25.62%of UPGR vs G_Baselineimprovement percentage -79.94% -76.59%of runing time vs G_Baselineto the traditional baseline G_Baseline, respectively, which fullydemonstrates the comprehensive effectiveness of TAHIN modulein insurance domain and cross domain recommendation method.Moreover, with the help of âĂŹagentâĂŹ heterogeneous auxiliaryinformation, the improvements of UPCR and UPGR in G_HCDIRare larger than that of G_HCDIR without-agent. As mentionedabove, agents are the key way to improve UPCR and UPGR in tradi-tional insurance recommendation, and it also proves that the strongpower of agent can significantly boost the performance of cold startproblem in online insurance recommendation. As for training time,the training time of G_Baseline model is 39.15 minutes, while thetraining time of G_HCDIR without agent and G_HCDIR are 7.86minutes and 9.17 minutes, respectively. As shown in Table 6, ourproposed HCDIR can at least improve by 76 %.
To our knowledge, there are not many papers about recommenda-tion systems in insurance products domain, some includes [6, 14,19, 20, 22]. [22] throughly describes the differences between recom-mendation system for classical domain and insurance domain, andfocuses on call centers servicing Life and Annual insurance, wherethe agents also have limited knowledge and experience. [6] proposea web recommendation system for life insurance sector by using as-sociation rules, which is one of the most well researched techniquesof data mining. [19] presents a hybird recommendation system ininsurance domain based on a standard user-user collaborate fil-tering approach. [20] utilizes Bayes networks to give customerspersonalized recommendation based on what other similar peoplewith similar portfolios have. [11] is a improved model of [20], whichtries to learn the structure of Bayesian network and considerablyspeeds up both training and inference run-times, while achievingsimilar accuracy. [14] propose a causation-driven visualization sys-tem that fundamentally transforms cross-media insurance data intonetwork diagrams and performs recommendation reasoning. How-ever, these methods neglect the item complexity and data sparsityproblem.
Cross-domain recommendation (CDR) [5, 12, 13, 16, 17, 28], whichaims to improve the recommendation performance by means oftransferring information from the auxiliary domain to the target domain, is one of the promising ways to solve data sparsity andcold start problem. Generally, CDR can be categorized into twocategories. One is to aggregate knowledge between two domains,this kind of methods are interested in improving the overall per-formance in the target domain [13, 16, 28], however, they can notdeal with cold start users. Since cold start users do not have anyinteractions in target domain. The other one aims at infering thepreferences of cold start users based on their preferences observedin other domains [5, 12, 17]. These methods assume that there existsoverlap in information between users and/or items across differentdomains, and train a mapping function from the source-domaininto the target-domain. For cold start users, these method first learnrepresentations in source domain, and then mapping them to thetarget domain.
Recently, some methods have been proposed representation learn-ing methods for HIN. These methods can be grossly divided intotwo groups: shallow models and deep models. Shallow models([3, 4, 9, 15] employ factorization-Âŋbased approaches or randomwalk approaches to aggregate information from neighbor nodes. Forexample, Metapath2vec [3] formalizes meta-paths based randomwalks to obtain heterogeneous neighborhoods of a node and lever-ages Skip-gram model to learn the network structure. However,this kind of method only explore one aspect information, failing tointegrate more heterogeneous information. Deep models [29, 31]aggregate neighbor information by neural network based method.HetGNN [31] jointly learn heterogeneous graph information andheterogeneous contents information for node embeddings based onGNN [23]. Inspired by graph attention networks, R-GCNs [24] aredeveloped to deal with the highly multi-relational data. HAN [29]designs a two level (node-level and semantic-level) attentions togenerate node embedding by aggregating features from meta-pathbased neighbors.
To deal with insurance product complexity and cold start problem,we propose a novel framework called a HCDIR for cold start usersin insurance domain. Specifically, we first try to learn more effectiveuser and item latent features in both source and target domains. Insource domain, we employ GRU to module users’ dynamic interests.
IGIR ’20, July 25–30, 2020, Virtual Event, China Ye Bi, Liqiang Song, Mengqiu Yao, Zhenyu Wu, Jianming Wang, Jing Xiao
In target domain, we construct an IHIN based on data from Jinguan-jia App, then we employ three-level (relational, node, and semantic)attention aggregations to get user and insurance product represen-tations. After obtaining the latent features of the overlapping users,a feature mapping between the two domains is learned by MLP.We apply HCDIR on PingAn Jinguanjia dataset, and show HCDIRsignificantly outperforms the state-of-the-art solutions. As futurework, we will try to construct more complete HIN, consideringmore types of relations, such as the relation between agent andinsurance product. We will also consider to train more accurateitem representations in source domain.
REFERENCES [1] Behnoush Abdollahi and Olfa Nasraoui. 2016. Explainable Matrix Factorizationfor Collaborative Filtering. In
Proceedings of the 25th International Conference onWorld Wide Web, WWW 2016, Montreal, Canada, April 11-15, 2016, CompanionVolume . 5–6. https://doi.org/10.1145/2872518.2889405 [2] Junyoung Chung, Çaglar Gülçehre, KyungHyun Cho, and Yoshua Bengio. 2014.Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Model-ing.
CoRR abs/1412.3555 (2014). arXiv:1412.3555 http://arxiv.org/abs/1412.3555 [3] Yuxiao Dong, Nitesh V. Chawla, and Ananthram Swami. 2017. metapath2vec:Scalable Representation Learning for Heterogeneous Networks. In
Proceedings ofthe 23rd ACM SIGKDD International Conference on Knowledge Discovery and DataMining, Halifax, NS, Canada, August 13 - 17, 2017 . 135–144. https://doi.org/10.1145/3097983.3098036 [4] Tao-Yang Fu, Wang-Chien Lee, and Zhen Lei. 2017. HIN2Vec: Explore Meta-paths in Heterogeneous Information Networks for Representation Learning.In
Proceedings of the 2017 ACM on Conference on Information and KnowledgeManagement, CIKM 2017, Singapore, November 06 - 10, 2017 . 1797–1806. https://doi.org/10.1145/3132847.3132953 [5] Wenjing Fu, Zhaohui Peng, Senzhang Wang, Yang Xu, and Jin Li. 2019. DeeplyFusing Reviews and Contents for Cold Start Users in Cross-Domain Recom-mendation Systems. In
The Thirty-Third AAAI Conference on Artificial Intelli-gence, AAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019 . 94–101. https://doi.org/10.1609/aaai.v33i01.330194 [6] Abdhesh Gupta and Anwiti Jain. 2013. Life insurance recommender system basedon association rule mining and dual clustering method for solving cold-startproblem.
International Journal of Advanced Research in Computer Science andSoftware Engineering
Advances in Neural Information ProcessingSystems 30: Annual Conference on Neural Information Processing Systems 2017, 4-9December 2017, Long Beach, CA, USA . 1024–1034.[8] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk.2016. Session-based Recommendations with Recurrent Neural Networks. In . http://arxiv.org/abs/1511.06939 [9] Binbin Hu, Chuan Shi, Wayne Xin Zhao, and Philip S. Yu. 2018. Leveraging Meta-path based Context for Top- N Recommendation with A Neural Co-AttentionModel. In Proceedings of the 24th ACM SIGKDD International Conference on Knowl-edge Discovery & Data Mining, KDD 2018, London, UK, August 19-23, 2018 . 1531–1540. https://doi.org/10.1145/3219819.3219965 [10] Binbin Hu, Zhiqiang Zhang, Chuan Shi, Jun Zhou, Xiaolong Li, and Yuan Qi.2019. Cash-Out User Detection Based on Attributed Heterogeneous InformationNetwork with a Hierarchical Attention Mechanism. In
The Thirty-Third AAAI Con-ference on Artificial Intelligence, AAAI 2019, Honolulu, Hawaii, USA, January 27 -February 1, 2019 . 946–953. https://doi.org/10.1609/aaai.v33i01.3301946 [11] Teja Kanchinadam, Maleeha Qazi, Joseph Bockhorst, Mary Y. Morell, Katie J.Meissner, and Glenn Fung. 2018. Using Discriminative Graphical Models for In-surance Recommender Systems. In .421–428. https://doi.org/10.1109/ICMLA.2018.00069 [12] SeongKu Kang, Junyoung Hwang, Dongha Lee, and Hwanjo Yu. 2019. Semi-Supervised Learning for Cross-Domain Recommendation to Cold-Start Users.In
Proceedings of the 28th ACM International Conference on Information andKnowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019 . 1563–1572. https://doi.org/10.1145/3357384.3357914 [13] Tzu-Heng Lin, Chen Gao, and Yong Li. 2019. CROSS: Cross-platform Recommen-dation for Social E-Commerce. In
Proceedings of the 42nd International ACM SIGIRConference on Research and Development in Information Retrieval, SIGIR 2019,Paris, France, July 21-25, 2019 . 515–524. https://doi.org/10.1145/3331184. 3331191 [14] Zhixiu Liu, Chengxi Zang, Kun Kuang, Hao Zou, Hu Zheng, and Peng Cui.2019. Causation-Driven Visualizations for Insurance Recommendation. In
IEEEInternational Conference on Multimedia & Expo Workshops, ICME Workshops 2019,Shanghai, China, July 8-12, 2019 . 471–476. https://doi.org/10.1109/ICMEW.2019.00087 [15] Yuanfu Lu, Chuan Shi, Linmei Hu, and Zhiyuan Liu. 2019. Relation Structure-Aware Heterogeneous Information Network Embedding. In
The Thirty-ThirdAAAI Conference on Artificial Intelligence, AAAI 2019, Honolulu, Hawaii, USA,January 27 - February 1, 2019 . 4456–4463. https://doi.org/10.1609/aaai.v33i01.33014456 [16] Muyang Ma, Pengjie Ren, Yujie Lin, Zhumin Chen, Jun Ma, and Maarten de Rijke.2019. π -Net: A Parallel Information-sharing Network for Shared-account Cross-domain Sequential Recommendations. In Proceedings of the 42nd InternationalACM SIGIR Conference on Research and Development in Information Retrieval,SIGIR 2019, Paris, France, July 21-25, 2019 . 685–694. https://doi.org/10.1145/3331184.3331200 [17] Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. 2017. Cross-DomainRecommendation: An Embedding and Mapping Approach. In
Proceedings of theTwenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI 2017,Melbourne, Australia, August 19-25, 2017 . 2464–2470. https://doi.org/10.24963/ijcai.2017/343 [18] Tomas Mikolov, Ilya Sutskever, Kai Chen, Gregory S. Corrado, and Jef-frey Dean. 2013. Distributed Representations of Words and Phrases andtheir Compositionality. In
Advances in Neural Information Processing Sys-tems 26: 27th Annual Conference on Neural Information Processing Sys-tems 2013. http://papers.nips.cc/paper/5021-distributed-representations-of-words-and-phrases-and-their-compositionality [19] Sanghamitra Mitra, Nilendra Chaudhari, and Bipin Patwardhan. 2014. Leveraginghybrid recommendation system in insurance domain.
International Journal ofEngineering and Computer Science
Proceedings ofthe Eleventh ACM Conference on Recommender Systems, RecSys 2017, Como, Italy,August 27-31, 2017 . 274–278. https://doi.org/10.1145/3109859.3109907 [21] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme.2012. BPR: Bayesian Personalized Ranking from Implicit Feedback.
CoRR abs/1205.2618 (2012). arXiv:1205.2618 http://arxiv.org/abs/1205.2618 [22] Lior Rokach, Guy Shani, Bracha Shapira, Eyal Chapnik, and Gali Siboni. 2013.Recommending insurance riders. In
Proceedings of the 28th Annual ACM Sym-posium on Applied Computing, SAC ’13, Coimbra, Portugal, March 18-22, 2013 .253–260. https://doi.org/10.1145/2480362.2480417 [23] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini. 2009. TheGraph Neural Network Model.
IEEE Transactions on Neural Networks
20, 1 (2009),61–80. https://doi.org/10.1109/TNN.2008.2005605 [24] Michael Sejr Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg,Ivan Titov, and Max Welling. 2018. Modeling Relational Data with GraphConvolutional Networks. In
The Semantic Web - 15th International Confer-ence, ESWC 2018, Heraklion, Crete, Greece, June 3-7, 2018, Proceedings . 593–607. https://doi.org/10.1007/978-3-319-93417-4_38 [25] Yizhou Sun, Jiawei Han, Xifeng Yan, Philip S. Yu, and Tianyi Wu. 2011. Path-Sim: Meta Path-Based Top-K Similarity Search in Heterogeneous InformationNetworks.
PVLDB
4, 11 (2011), 992–1003.[26] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, PietroLiò, and Yoshua Bengio. 2018. Graph Attention Networks. In
Proceesings of the6th International Conference on Learning Representations, ICLR 2018, Vancouver,BC, Canada, April 30 - May 3, 2018 . https://openreview.net/forum?id=rJXMpikCZ [27] Minjie Wang, Lingfan Yu, Da Zheng, Quan Gan, Yu Gai, Zihao Ye, Mufei Li,Jinjing Zhou, Qi Huang, Chao Ma, Ziyue Huang, Qipeng Guo, Hao Zhang, HaibinLin, Junbo Zhao, Jinyang Li, Alexander J Smola, and Zheng Zhang. 2019. DeepGraph Library: Towards Efficient and Scalable Deep Learning on Graphs. ICLRWorkshop on Representation Learning on Graphs and Manifolds (2019). https://arxiv.org/abs/1909.01315 [28] Xiang Wang, Xiangnan He, Liqiang Nie, and Tat-Seng Chua. 2017. Item Silk Road:Recommending Items from Information Domains to Social Users. In
Proceedingsof the 40th International ACM SIGIR Conference on Research and Developmentin Information Retrieval, Shinjuku, Tokyo, Japan, August 7-11, 2017 . 185–194. https://doi.org/10.1145/3077136.3080771 [29] Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S.Yu. 2019. Heterogeneous Graph Attention Network. In
The World Wide WebConference, WWW 2019, San Francisco, CA, USA, May 13-17, 2019 . 2022–2032. https://doi.org/10.1145/3308558.3313562 [30] Fengli Xu, Jianxun Lian, Zhenyu Han, Yong Li, Yujian Xu, and Xing Xie. 2019.Relation-Aware Graph Convolutional Networks for Agent-Initiated Social E-Commerce Recommendation. In
Proceedings of the 28th ACM International Con-ference on Information and Knowledge Management, CIKM 2019, Beijing, China,November 3-7, 2019 . 529–538. https://doi.org/10.1145/3357384.3357924
Heterogeneous Information Network based Cross Domain Insurance Recommendation System for Cold Start Users SIGIR ’20, July 25–30, 2020, Virtual Event, China [31] Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V.Chawla. 2019. Heterogeneous Graph Neural Network. In