DEVELOPING PRICING MODELS FOR ONLINE ART SALES USING

DEVELOPING PRICING MODELS FOR ONLINE ART SALES USING TEXT ANALYTICS LAUREL POWELL, ZBIGNIEW W. RAS COMPUTER SCIENCE DEPT. , COLLEGE OF COMPUTING & INFORMATICS (CCI) SA UNIVERSITY OF NORTH CAROLINA, CHARLOTTE, USA & ANNA GELICH CS SCHOOL OF ARCHITECTURE, COLLEGE OF ARTS + ARCHITECTURE UNIVERSITY OF NORTH CAROLINA, CHARLOTTE, USA Research Supported by

OUTLINE • Introduction • Current Art Market • Pricing in the Art Market (galleries, online sales, street market, …) • Features having impact on Price (size, name of artist, medium…. ) • Knowledge Based Recommender System for Assigning Price Tags to Art • Current results

INFORMATION ABOUT PAINTING/ARTIST – AN EXAMPLE • Anastasia Kachina • Size: 11. 81 x 15. 75 in • $405 -> sold for $350 by Saatchi Art about 1 year • 431 artworks ago • Education: MA • Kharkiv Art College (1998 -2003) • Kharkiv State Academy of Design and Arts (20032010) • Exhibitions: 10 • Locations and Type of Exhibitions The same type of her painting costs now about $700 (one year ago she had much less artworks)

INFORMATION ABOUT PAINTING/ARTIST – ANOTHER EXAMPLE • Annabel Andrews • $5, 200 – asking price by Saatchi Art • Born in 1940, lives in Spain • Artworks - 154 • Many individual and collective exhibitions • Received many prizes Asking price - $1, 300

CONTEMPORARY ART MARKET • Multi-billion dollar industry • 63. 7 Billion USD in 2017 • More galleries closed than opened (what business model to follow to be successful? ) NPS based or GET-KEEP-GROW? • Highly uncertain, and has proven to be a challenge to explain using traditional economic theories. What about data analytics? • 29% of sales take place online (Saatchi. Art, Artfinder, …. ) • Price transparency key concern of online buyers

ART AND RECOMMENDER SYSTEMS • Existing recommender systems in the art market domain do not use data analytics but human experts to evaluate fine art pieces and make recommendations. • Mutual. Art (https: //www. mutualart. com/artappraisal) has the world’s most comprehensive database of past sale results but the number of features describing these sales is quite small. Its advisors are assigning price tags to new pieces of art by comparing them with similar pieces in the Mutual. Art database. • The charge for a single service (one piece of art) is $49 and the waiting time to get a recommendation is 72 hours

ART AND RECOMMENDER SYSTEMS • FINDARTINFO (http: //www. findartinfo. com/english. html) is a similar but free art appraisal service which contains information about 438, 003 artists and 3, 775, 762 art prices. With this art appraisal tool, an artist can value his/her fine art by comparing it with recent auction prices of similar pieces. There are websites providing free art appraisal hints. • For instance, wiki. How (https: //www. wikihow. com/Value-Your-Art) helps to value artworks.

RECOMMENDER SYSTEM - ARTIST • It is based on data analytics. Knowing the artist’s name, appraisal of the piece of art is done by Art. IST by its personalized discovery module built from the data describing similar artists and similar art pieces including their sales. • Dataset • Base Features (medium, size of painting, subject, style, artist name, features describing artist reputation, …) • New Features Developed and Tested • Word Count (in artist & artworks provided information) • Social Media (presence or absence of social media links in artist provided information) • Clustering & sentiment analysis of artworks provided information

• Scraped from Artfinder. com • About 200, 000 artworks (currently more than ½ million) • 2, 800 artists (currently more than 5, 000 artists) • 60 countries DATASET • Artist Pages (information about artists which has optional sections: CV, education, awards received, links to other locations presenting the artist, …) • Artwork Pages • Reviews We collaborate with QCollector which has data about 5 million paintings being sold by them.

PRICE DISTRIBUTIO N Maximum 1, 000 USD Minimum 12. 97 USD Figure omits works priced at greater than 5000 USD due to scale

BASE FEATURES • – artist. ID - A unique identifier for an artist • – artist. Country - The artist’s current country of residence as listed on their profile • – percent n stars (n=1, 2, 3, 4, 5) - Percentage of n star reviews out of the total number of reviews. • – medium - Artist provided medium of the artwork. • – style - Artist provided style of the artwork. • – subject - Artist provided subject of the artwork. • – authentication - Artist provided method of authenticating the work. • – artwork width - The width of the artwork. • – artwork height - The height of the artwork.

DISCRETIZATION • Price dimension was reduced to discrete intervals. • These intervals are as follows: (0 - 105), [105 - 205), [205 - 405), [405 - 605), [605 - 810), [810 - 1030), [1030 1445), [1445 - 1825), [1825 - 2455), [2455 - 3855), [3855 5000), [5000 - 10, 000), (> 10, 000). The intervals were selected by searching for areas where very few works were priced and splitting the price feature in these gaps.

Confusion Matrix with Base Features for Artfinder Dataset

RESULTS WITH BASE SET OF FEATURES Method F 1 Score Precision Recall k. NN 0. 621 0. 619 0. 625 SVM 0. 17 0. 221 0. 169 Random Forest 0. 663 0. 662 0. 667

RESULTS WITH NEW FEATURES: BIOGRAPHY, DESCRIPTION OF ARTWORKS, TITLE

RESULTS WITH SOCIAL MEDIA PRESENCE

TEXT CLUSTERING Awards Biography Descriptio n Education Events Title

TEXT CLUSTERING • Paragraph Vector • Also known as Doc 2 Vec • Extension of Word 2 Vec • 10, 25 and 50 Clusters • Biography, Awards, Events and Education most impactful We are currently replacing Paragraph Vector by Folksonomy

RESULTS WITH 10 CLUSTERS

RESULTS WITH 25 CLUSTERS

RESULTS WITH 50 CLUSTERS

ALL FEATURES (25 CLUSTERS)

• VADER SENTIMENT ANALYSIS • Valence Aware Dictionary for s. Entiment Reasoning • Rules and Lexicon • Biography, Description of Artworks, Title

COMBINED FEATURES F 1 Precision Recall Base Features 0. 664 0. 663 0. 667 BF, Social, Word Counts 0. 677 0. 676 0. 681 BF, Social, Word Counts, Clusters (10) 0. 674 0. 678 BF, Social, Word Counts, Clusters (25) 0. 676 0. 675 0. 68 BF, Social, Word Counts, Clusters (50) 0. 679 0. 678 0. 683 BF, Social, Word Counts, Sentiment 0. 685 0. 684 0. 689 BF, Social, Word Counts, Sentiment, Clusters (50) 0. 685 0. 69

With Base Features Confusion Matrix with Base Features and All Added Features