[추천시스템] E-Commerce Product Recommendation Agents : Use, Characterstics, and Impact

◈ 논문 정보 

 

• Xiao, B., & Benbasat, I. (2007). E-commerce product recommendation agents: Use, characteristics, and impact. MIS quarterly, 137-209.

 

---

◈ 요약 

 본 연구에서는 RA(Recommendation Agents)의 효과에 영향을 미치는 중요 요소들을 통합하여 conceptual model을 제시했다. 선행 연구 사례에서 총 28개의 특징을 도출하고 이를 TAM 모델처럼 하나의 큰 모델로 정립하고자 하였다. 

 

RQ 1. How do RA use, RA characterstics, and other factors influence consumer decision making processes and outcomes? 

1.1 RA의 사용은 유저들의 의사결정과 결과에 어떻게 영향을 미칠 것인가? 

1.2 RA의 특징은 유저들의 의사결정과 결과에 어떤 영향을 미칠 것인가? 

1.3 RA와 관련된 다른 요소들은, 1.1과 1.2에 대해서 어떤 조절 효과를 보일 것인가? 

 

RQ 2. How do RA use, RA characterstics, and other factors influence users' evaluation of RAs? 

2.1 RA사용은 유저들의 RA 평가에 어떤 영향을 미칠 것인가?

2.2 RA의 특징은 RA에 대한 유저들의 평가에 어떤 영향을 미칠 것인가?

2.3 RA와 관련된 다른 요소들은 2.1과 2.2에 대해서 어떤 조절효과를 보일 것인가? 

2.4 provider credibility는 유저들의 RA 평가에 어떤 영향을 미칠 것인가. 

 

 

 연구 모델은 위와 같으며, 천시스템과 관련이 있는 전반적인 요소들을 다양하게 모델에 포함시켰다. 이 연구에서는 outcome variable로 크게 두 가지 요소를 고려하였다. 첫 번째는 Consumer Decision Making인데, 구체적으로 RA의 사용 여부와 RA의 특징에 따른 차이를 살펴보았다. 두 번째는 RA에 대한 인식이다. 이는 user's evaluation of RA 항목으로 측정된다. 구체적인 항목으로는 TAM 모델의 요소들과 만족도(satisfaction)를 사용했다. 또한 선행 연구의 결과를 종합해서 RA characteristics, User, User-Ra Interaction, Product, Provider credibility 항목들은 조절 변수로 추가하였다. 독립 변수로는 RA Use를 사용했다. 

 

 연구 가설은 총 28개이며 하위 가설을 포함하면 40개가 넘는다. 지금까지 본 연구 중에서 모델이 가장 복잡하고 가설의 수도 가장 많았다. 

P1~P13은 RQ 1과 관련이 있으며 human information processing theory와 theory of interpersonal similarity 를 기반으로 한다. 

 

DE(Decision Effort) 

DQ(Decision Quality) 

 

• P1 : RA use influences user's decision effort 

  P1a : RA use reduces the extent of product search 

  P1b : RA use reduces user's decision time 

  P1c : RA use increases the amount of user input 

 

• P2 : RA use improves users' decision effort(DE) and decision quality (DQ)

 

• P3 : RA type influences users' decision effort and decision quality 

  P3a : Compared with pure content-filtering or CF RAs, hybrid RAs lead to better DQ and higher DE

  P3b : Compared with non-compensatory RAs, compensatory RAs lead to better DQ and higher DE

  P3c : Compared with feature-based RAs, needs-based RAs lead to better DQ

 

• P4 : Preference elicitation method influences users's DQ and DE. The explicit preference elicitation method lead to better DQ and higher DE thatn implicit preference elicitation method. 

 

• P5 : Included product attributes influences users' preference function and choice. 

        They are given more weight in users' preference function and considered more important. 

 

• P6 : Recommendation content influences users' product evaluation and choice. 

  P6a : Recommendation provided by RAs influence users' choice to the extent that products recommended RAs are more likely to be chosen by users. 

  P6b : The display of utility score or predicted ratings for recommended products influences users' product evaluation and choice. 

  

• P7 : Recommendation format influences users' decision process and decision outcome.

  P7a : Recommendation display method influences users' decision strategies and DQ to the extent that sorted recommendation lists result in greater user reliance. 

  P7b : The number of recommendations influences users' DE and DQ to the extent that presenting too many recommendations increases DE and decreases DQ.

 

• P8 : Product type moderates the effects of RA use on users' choice.

  RA use influences users shopping for experience products to greater extent than search products. 

 

• P9 : Product complexity moderated the effects of RA use on users' DQ and DE. 

 Use of RAs for more complex products lead to greated increase in DQ and decrease in DE. 

 

• P10 : Product complexity moderates the effect of included product attributes on users' choice. 

 The inclusion effect is stronger for products with negative inter-attribute correlations thatn for those with positive inter-attribute correlations. 

 

• P11 : Product expertise moderates the effect of preference elicitation method on users' DQ. 

  Preference elicitation method has less effect of DQ of users with high product expertise. 

 

• P12 : Perceived product risks moderates effects of RA use on users' DQ and DE. 

  When perceived high proudct risks, RA use lead to greated improvements in DQ and reduction in DE. 

 

• P13 : User-RA similarity moderates the effect of RA use on users' DQ and DE. 

  RA use leads to increase in DQ and decrease DE when RAs are similar to the users. 

  (여기서 similarity란, RA가 얼마나 유저의 취향을 잘 반영하여 유사하게 추천하는가를 의미한다) 

 

 

 

P14~P28은 RQ 2.와 관련된 가설들이며 다섯 가지 이론을 기반으로 한다. 

 

PU : Perceived Usefulness

PEU : Perceived Ease of Use 

 

• P14 : RA type influences users' evaluation of RAs.

  P14a : Compared with pure content-filtering or pure CF RAs, hybrid RAs lead to greater trust, PU, satisfaction but lower PEU. 

 P14b : Compared with non-compensatory RAs, compensatory RAs lead to greater turst, PE, satisfaction but lower PEU.

 

• P15 : The preference elicitation method influences users' PEU and satisfaction with RAs. 

  Compared to an explicit preference elicitation method, implicit preference method leads to greater PEU and satisfaction. 

 

• P16  : The ease of generating new of additional recommendations influences users' PEU and satisfaction with RAs. 

  The easier it is for the users to generate new recommendations, the greater their PEU and satisfaction. 

 

• P17 : User control of interaction with RA's preference-elicitation interface influences users' trunt in, satisfaction with, and PU of RAs. Increased user control lead to increased trust, satisfaction and perception of usefulness. 

 

• P18 : The provision of information about search progress, while users await results influences users' satisfaction with RAs. Users who are informed about serach progress are more satisfied with the RAs. 

 

• P19 : Response time influences users' satisfaction with RAs. 

  The longer the response times, the lower the users' satisfaction with the RAs. 

 

• P20 : Recommendation content influences users' evaluations of RAs

  P20a : Familiar recommendations increase users' trust in the RAs

  P20b : The composition of the recommendation list, as reflected by balanced representation of familiar and unfamiliar product recommendation, influences users' trust in PU and satisfaction. 

  P20c : The provision of detailed information about recommended products increases users' trust, PU and satisfaction with RAs. 

  P20d : The provision of explanation on how the recommendations are generated increases users' trust and satisfaction with RAs. 

 

• P21 : Recommendation format influences users' PU, PEU, satisfaction with RAs. RAs with clear navigational path and layout are considered more useful. 

 

• P22 : Product type moderates the effects of RA use on users' trust and PU of RAs. 

  Users have higher trust for experience products and higher PU for search products. 

 

• P23 : Product expertise moderates the effect of RA use on users' evaluations of RAs. 

   The higher the product expertise of the users, the less favorable the evaluation of RAs. 

 

• P24 : Product expertise moderates the effects of RA type on users' evaluations of RAs. 

  The higher the product expertise of users, the more favorable the users' evaluation of feature-based RAs. 

  Also, the more favorable the user's evaluations of contetn-filtering RAs. 

 

• P25 : User–A similarity moderates the effects of RA use on users’ trust in, satisfaction with, and perceived usefulness of RAs.

The more the RAs are perceived to be similar to their users, the more they are considered to be trustworthy, satisfactory, and useful.

 

• P26 : User’s familiarity with RAs moderates the effects of RA use on trust in the RAs. Increased familiarity with RAs
leads to increased trust in the RAs.

 

• P2& : The confirmation/disconfirmation of expectations about RAs moderates the effects of RA use on users’ satisfaction with the RAs.

 Confirmation or positive disconfirmation of users’ expectations about RAs contributes positively to users’ satisfaction with the RAs. In contrast, negative disconfirmation of users’ expectations about RAs contributes negatively to users’ satisfaction
with the RAs.

 

• P28 : Provider credibility, determined by the type of RA providers and the reputation of RA providers, influences users’ trust in RAs.

 RAs provided by independent third party websites are considered more trustworthy than those provided by vendors’ websites. RAs provided by reputable websites are considered more trustworthy than those provided by websites that are unknown or non-
reputable.

 

 

 각 proposition에 대한 결과를 보면 대부분이 지지되었음을 알 수 있다. 또한 일부 가설들은 실증적으로 검증되지 않은 것을 확인할 수 있다. 

 

 

 

 

---

◈ 장점 및 의의 

• 당시 시점을 기준으로 추천시스템과 관련된 여러 연구의 내용들을 통합하여 하나의 체계로 정리하였다. 추천시스템 내외부의 다양한 요소들이 어떻게 상호 관련이 있는지 일목요연하게 확인할 수 있었다. 

 

• 본문에서 제시한 각 proposition에 대해서 합리적인 이론적 근거들을 제시하였다. 그래서 이 논문이 단순히 기존 연구 결과들을 정리한 것 이상의 가치를 지닌다고 생각한다. 

 

 

 

---

◈ 한계점 및 추가 연구 아이디어 

• Suggestion for future research 부분의 내용들을 참고해 볼 만 한다. 물론 지금 시점으로는 관련된 많은 연구가 진행되었을 것이다.