EvidencefromAmazon.com Personality-BasedRecommendations...
1
Personality-Based Recommendations: Empirical Evidence from Amazon.com Panagiotis Adamopoulos and Vilma Todri Leonard N. Stern School of Business, New York University {padamopo,vtodri}@stern.nyu.edu A BSTRACT ◦ We evaluate the accuracy of personality-based rec- ommendations using a real-world data set from Amazon.com. ◦ We automatically infer the personality traits, needs, and values of users based on unstructured user- generated content in social media. ◦ Personality characteristics significantly increase the performance of recommender systems. ◦ Different personality models exhibit statistically sig- nificant differences in predictive performance. M OTIVATION ◦ Personality traits have been found to influence various aspects of individual behavior. ◦ Personality traits are still not frequently used in pre- dictive modeling, mainly because they usually require users to complete long questionnaires and, hence, they cannot be easily applied at large scale. ◦ Abundance of user-generated text in social media leads to opportunities to automatically derive personality traits based on the established relationship between word use and personality. ◦ Multiple personality models, including Big Five, Needs, and Values. P ERSONALITY I NFERENCE The personality traits [3,6], needs [4,5], and values [7] of the users are automatically inferred based on a textual analysis of unstructured user-generated data from social media: ◦ For each user we analyzed the content of all the mes- sages that there were publicly posted over time on the social network of Twitter. ◦ After the pre-processing of the corpus of user- generated content, there were on average 26, 568 words per user. ◦ The tokens of the user-generated content are matched with the Linguistic Inquiry and Word Count (LIWC) psycholinguistic dictionary to compute relative scores in each dictionary category. ◦ A weighted combination is estimated based on the cor- relation between category scores and personality char- acteristics. DATA ◦ Similar to [1, 8], we collect a real-world data set based on publicly available and well-structured tweets refer- ring to online purchases from Amazon.com. ◦ This data set is collected and analyzed in real time us- ing the Twitter streaming API. ◦ Our data set contains information on: ◦ 81, 475 distinct Amazon customers, who shared ◦ 906, 277 purchases on Twitter.com, and ◦ 138, 536 unique products. R ECOMMENDATION M ODEL To empirically evaluate the employed approach, we build a factorization model incorporating the information of per- sonality traits, needs, and values as well as item attributes. In particular, the user preferences are modeled as: y (x)= y (u; i; α u 1 ,...,α u m ; α i 1 ,...,α i n ) = w 0 + w u + w i + m X j =1 w j α u j + n X l=1 w l α i l + hv u , v i i + m X j =1 α u j hv u j , v i i + n X l=1 α i l hv u , v i l i + m X j =1 n X l=1 α u j α i l hv u j , v i l i, where the input vector x ∈ R |U |+|I |+m+n contains binary indicators for the user and item, the user attributes α u = (α u 1 ,...,α u m ) capturing the personality characteristics of the users, and item attributes α i =(α i 1 ,...,α i m ) capturing the item categories, prices, etc.; the factorization of users v u , items v i , and attributes v u j , v i l is of dimensionality k . E XPERIMENTAL S ETTING ◦ For the various specifications of the factorization model we use: i) the basic factorization model, which also includes item attributes such as prices and product categories, ii) the personality traits of the users based on the Big Five personality model [3, 6], iii) the twelve categories of needs of the users as identified in marketing literature [4, 5], and iv) the basic value priorities of the users [7]. ◦ For each user we randomly select an equal number of non-rated items as negative examples; these negative examples are sampled based on the frequency of purchase of each item. ◦ We use MCMC inference with Gibbs sampling to learn our factorization model. ◦ We employ a holdout evaluation scheme with 80/20 random splits into training and test sets. ◦ We evaluate each model in term of classification performance based on accuracy. E XPERIMENTAL R ESULTS Figure 1: Predictive performance of different personality models. Based on the experimental results presented in Figure 1, we see that: ◦ personality characteristics increase the performance of recommender systems, ◦ different personality models can result in different predictive accuracy, ◦ the under-explored personality models of needs [4, 5] and values [7] resulted in better predictive performance compared to the more popular model of Big Five traits [3, 6], and that ◦ combining the attributes of the different personality models results in even better performance. F UTURE D IRECTIONS The proposed approach should be extended to the task of rating prediction. Besides, future research should estimate the economic value of different personality traits and models in various recommendation domains and settings [2]. R EFERENCES [1] Panagiotis Adamopoulos. ConcertTweets: A Multi-Dimensional Data Set for Recommender Systems Research, 2014. [2] Panagiotis Adamopoulos and Alexander Tuzhilin. Estimating the Value of Multi-Dimensional Data Sets in Context-based Recommender Systems. In RecSys ’14, 2014. [3] Paul Costa and Robert MacCrae. Revised NEO Personality Inventory (NEO PI-R) and NEO Five-factor Inventory (NEO FFI) Manual. Psychological Assessment Resources, 1992. [4] Kevin Ford. Brands Laid Bare: Using Market Research for Evidence-based Brand Management. John Wiley & Sons, 2005. [5] Philip Kotler and Gary Armstrong. Principles of Marketing. Pearson, 2013. [6] Warren Norman. Toward an Adequate Taxonomy of Personality Attributes: Replicated Factor Structure in Peer Nomination Personality Ratings. The Journal of Abnormal and Social Psychology, 66(6), 1963. [7] Shalom Schwartz. Basic Human Values: Theory, Measurement, and Applications. Revue française de sociologie, 47(4), 2006. [8] Vilma Todri and Panagiotis Adamopoulos. Social commerce: An empirical examination of the antecedents and consequences of commerce in social network platforms. In ICIS, 2014.
Transcript of EvidencefromAmazon.com Personality-BasedRecommendations...
Personality-Based Recommendations: EmpiricalEvidence from Amazon.com
Panagiotis Adamopoulos and Vilma TodriLeonard N. Stern School of Business, New York University
{padamopo,vtodri}@stern.nyu.edu
ABSTRACT◦ We evaluate the accuracy of personality-based rec-
ommendations using a real-world data set fromAmazon.com.
◦ We automatically infer the personality traits, needs,and values of users based on unstructured user-generated content in social media.
◦ Personality characteristics significantly increase theperformance of recommender systems.
◦ Different personality models exhibit statistically sig-nificant differences in predictive performance.
MOTIVATION◦ Personality traits have been found to influence various
aspects of individual behavior.◦ Personality traits are still not frequently used in pre-
dictive modeling, mainly because they usually requireusers to complete long questionnaires and, hence, theycannot be easily applied at large scale.
◦ Abundance of user-generated text in social media leadsto opportunities to automatically derive personalitytraits based on the established relationship betweenword use and personality.
◦ Multiple personality models, including Big Five,Needs, and Values.
PERSONALITY INFERENCEThe personality traits [3,6], needs [4,5], and values [7] ofthe users are automatically inferred based on a textualanalysis of unstructured user-generated data from socialmedia:◦ For each user we analyzed the content of all the mes-
sages that there were publicly posted over time on thesocial network of Twitter.
◦ After the pre-processing of the corpus of user-generated content, there were on average 26, 568 wordsper user.
◦ The tokens of the user-generated content are matchedwith the Linguistic Inquiry and Word Count (LIWC)psycholinguistic dictionary to compute relative scoresin each dictionary category.
◦ A weighted combination is estimated based on the cor-relation between category scores and personality char-acteristics.
DATA◦ Similar to [1, 8], we collect a real-world data set based
on publicly available and well-structured tweets refer-ring to online purchases from Amazon.com.
◦ This data set is collected and analyzed in real time us-ing the Twitter streaming API.
◦ Our data set contains information on:
◦ 81, 475 distinct Amazon customers, who shared◦ 906, 277 purchases on Twitter.com, and◦ 138, 536 unique products.
RECOMMENDATION MODELTo empirically evaluate the employed approach, we build a factorization model incorporating the information of per-sonality traits, needs, and values as well as item attributes. In particular, the user preferences are modeled as:
y(x) = y(u; i;αu1 , . . . , α
um;αi
1, . . . , αin)
= w0 + wu + wi +m∑j=1
wjαuj +
n∑l=1
wlαil + 〈vu,vi〉
+m∑j=1
αuj 〈vu
j ,vi〉+n∑
l=1
αil〈vu,v
il〉+
m∑j=1
n∑l=1
αuj α
il〈vu
j ,vil〉,
where the input vector x ∈ R|U |+|I|+m+n contains binary indicators for the user and item, the user attributes αu =(αu
1 , . . . , αum) capturing the personality characteristics of the users, and item attributes αi = (αi
1, . . . , αim) capturing
the item categories, prices, etc.; the factorization of users vu, items vi, and attributes vuj ,v
il is of dimensionality k.
EXPERIMENTAL SETTING◦ For the various specifications of the factorization model we use:
i) the basic factorization model, which also includes item attributes such as prices and product categories,ii) the personality traits of the users based on the Big Five personality model [3, 6],
iii) the twelve categories of needs of the users as identified in marketing literature [4, 5], andiv) the basic value priorities of the users [7].
◦ For each user we randomly select an equal number of non-rated items as negative examples; these negativeexamples are sampled based on the frequency of purchase of each item.◦ We use MCMC inference with Gibbs sampling to learn our factorization model.◦ We employ a holdout evaluation scheme with 80/20 random splits into training and test sets.◦ We evaluate each model in term of classification performance based on accuracy.
EXPERIMENTAL RESULTS
Figure 1: Predictive performance of different personality models.
Based on the experimental results presented in Figure 1, we see that:◦ personality characteristics increase the performance of recommender systems,◦ different personality models can result in different predictive accuracy,◦ the under-explored personality models of needs [4, 5] and values [7] resulted in better predictive
performance compared to the more popular model of Big Five traits [3, 6], and that◦ combining the attributes of the different personality models results in even better performance.
FUTURE DIRECTIONSThe proposed approach should be extended to the task of rating prediction. Besides, future research should estimate theeconomic value of different personality traits and models in various recommendation domains and settings [2].
REFERENCES
[1] Panagiotis Adamopoulos. ConcertTweets: A Multi-Dimensional Data Set for Recommender Systems Research, 2014.[2] Panagiotis Adamopoulos and Alexander Tuzhilin. Estimating the Value of Multi-Dimensional Data Sets in Context-based
Recommender Systems. In RecSys ’14, 2014.[3] Paul Costa and Robert MacCrae. Revised NEO Personality Inventory (NEO PI-R) and NEO Five-factor Inventory (NEO FFI)
Manual. Psychological Assessment Resources, 1992.[4] Kevin Ford. Brands Laid Bare: Using Market Research for Evidence-based Brand Management. John Wiley & Sons, 2005.[5] Philip Kotler and Gary Armstrong. Principles of Marketing. Pearson, 2013.[6] Warren Norman. Toward an Adequate Taxonomy of Personality Attributes: Replicated Factor Structure in Peer Nomination
Personality Ratings. The Journal of Abnormal and Social Psychology, 66(6), 1963.[7] Shalom Schwartz. Basic Human Values: Theory, Measurement, and Applications. Revue française de sociologie, 47(4), 2006.[8] Vilma Todri and Panagiotis Adamopoulos. Social commerce: An empirical examination of the antecedents and consequences of
commerce in social network platforms. In ICIS, 2014.
