Hinnaelastsusel tuginev soovitussüsteem
Date
2016
Authors
Pajula, Gea
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Soovitussüsteeme on palju uuritud ja edukalt rakendatud paljudes valdkondades, et suurendada läbimüüki tehes klienditele asjakohaseid soovitusi. Käesoleva magistritöö eesmärgiks on välja töötada uudne soovitussüsteem, mis teeb klientidele personaalseid pakkumisi toote soodushinna huvipakkuvuse põhjal. Seda saab rakendada olukordades, kus soovitusi tehakse allahinnatud toodete seast. Näiteks valides kampaaniatooteid kliendile saadetavasse personaalsesse uudiskirja. Me kaasame tootepõhise kaasfiltreerimise algoritmi täiendusena majanduse valdkonnas kasutatavat nõudluse hinnaelastsust, et võtta arvesse, et tootel on kampaaniaperioodil tavapärasest madalam hind. Hinnates mudeli abil omaelastsuse väärtuse, saame kliendi tootereitingu, mis näitab, kuidas hinna muutumine mõjutab ostetavat kogust. Toodete sarnasuste leidmiseks kasutame ristelastsust, mis liigitab tooted asendus- ja täiendkaupadeks. Selle suuruse leidmine ei nõua tihti esinevat kaugusmõõtude tingimust, et kaks toodet peavad olema ostetud samade klientide poolt. Kirjeldatud soovitussüsteem on rakendatud reaalelulistele supermarketi tehingute andmetele. Süsteemi headuse testimiseks kasutame kahte kampaaniaperioodi, mille allahinnatuid tooteid kasutame võimalike soovitustena. Me saavutame märgatavalt paremad tulemused kasutades ainult kampaaniatoote elastsusi ja mitte asendustoodete vastavaid väärtusi. Kaasates ainult kliendid, kellele leidsime vähemalt 5 pakkumist, saavutame tunduvalt paremad tulemused. Täpsemalt, tehes neile klientidele 12 soovitust (vähem kui 1% kampaaniatoodete arvust), tabame kõik klientide kampaaniatoodete ostud. Parima meetodi korral saavutame kordustäpsuse 0,24, mis on üle 10 korra parem võrreldes meetodiga, mida ettevõte hetkel kasutab, kus soovitused on manuaalselt valitud kliendi segmentide omaduste põhjal. Supermarketi kett on kinnitanud oma soovi, et esitletud meetodit testida, seega antud soovitussüsteemi rakendatakse reaalselt klientidele huvipakkuvate soovituste tegemiseks.
Recommender systems have been widely studied and successfully applied in a variety of areas to increase sales by guiding people toward items they are more likely to find interesting. The aim of this thesis is to develop a novel recommender system that suggests items to a client based on the appeal of a product discount. This can be applied to situations where recommendations are made from a list of discounted items such as campaign products selected into personalized sales promotion letters. To take into consideration that the products have cheaper price than usual during the campaign period, we propose an extension to an item based collaborative filtering algorithm, namely the price elasticity of demand known from the field of economics. We represent a client's rating about an item by estimating with a model own elasticity which measures the sensitivity of quantity demanded to the changes in the prices. The similarities of items are computed using cross elasticity which exhibits the substitutional and complementary effects among the products. Unlike traditional similarity metrics, this measure does not assume that the two items have to be purchased by the same clients. The proposed recommender system based on the price elasticity is applied to a real world supermarket transactions dataset. The performance of the system is evaluated on two campaign periods where recommendations are made from the discounted products. The experiments show that it is better to make recommendations based on only campaign product elasticities, without considering the elasticities of campaign products' substitutes. Furthermore, when including only the customers for whom we have found at least 5 recommendations, the performance is considerably better. In particular, when making 10 suggestions (less than 1% of all campaign products), we detect all campaign products that the clients indeed purchased. Using the best method, our approach achieves precision of 0.24, which is over 10 times better in comparison to the method currently used by the company where employees manually select recommendations based on the characteristics of customer segments. The supermarket chain has confirmed their interest in testing the proposed method in practice, hence it will be applied in real world to make more relevant recommendations to the customers.
Recommender systems have been widely studied and successfully applied in a variety of areas to increase sales by guiding people toward items they are more likely to find interesting. The aim of this thesis is to develop a novel recommender system that suggests items to a client based on the appeal of a product discount. This can be applied to situations where recommendations are made from a list of discounted items such as campaign products selected into personalized sales promotion letters. To take into consideration that the products have cheaper price than usual during the campaign period, we propose an extension to an item based collaborative filtering algorithm, namely the price elasticity of demand known from the field of economics. We represent a client's rating about an item by estimating with a model own elasticity which measures the sensitivity of quantity demanded to the changes in the prices. The similarities of items are computed using cross elasticity which exhibits the substitutional and complementary effects among the products. Unlike traditional similarity metrics, this measure does not assume that the two items have to be purchased by the same clients. The proposed recommender system based on the price elasticity is applied to a real world supermarket transactions dataset. The performance of the system is evaluated on two campaign periods where recommendations are made from the discounted products. The experiments show that it is better to make recommendations based on only campaign product elasticities, without considering the elasticities of campaign products' substitutes. Furthermore, when including only the customers for whom we have found at least 5 recommendations, the performance is considerably better. In particular, when making 10 suggestions (less than 1% of all campaign products), we detect all campaign products that the clients indeed purchased. Using the best method, our approach achieves precision of 0.24, which is over 10 times better in comparison to the method currently used by the company where employees manually select recommendations based on the characteristics of customer segments. The supermarket chain has confirmed their interest in testing the proposed method in practice, hence it will be applied in real world to make more relevant recommendations to the customers.