Hide
Раскрыть

ISSN 2587-814X (print),
ISSN 2587-8158 (online)

Russian version: ISSN 1998-0663 (print),
ISSN 2587-8166 (online)

Alexander D’yakonov1
  • 1 Lomonosov Moscow State University; Dorodnitsyn Computing Center, Russian Academy of Sciences , 1, build. 52, Lomonosov Moscow State University, Leninskie Gory, GSP-1, Moscow, 119991, Russian Federation

Supermarkets clients behaviour forecasting by weighted methods of probability and density estimations

2014. No. 1 (27). P. 68–77 [issue contents]

Alexander D’yakonov - Professor, Department of Mathematical Methods of Forecasting, Faculty of Computational Mathematics and Cybernetics, Lomonosov Moscow State University; Senior Researcher, Dorodnitsyn Computing Center, Russian Academy of Sciences  
Address: 1, build. 52, Lomonosov Moscow State University, Leninskie Gory, GSP-1, Moscow, 119991, Russian Federation.
E-mail: djakonov@mail.ru

     We consider two tasks in describing a supermarkets clients’ behavior: prediction of a client's next visit date and prediction of his/her spends. The first problem is equal to estimating visit probability, and the second – to estimating density for visitor spends. To solve these problems, we propose using weighed methods: real non-negative value (weight) is assigned to every event. Weights allow considering additional information, for example history (earlier visits have smaller weights). We consider several weighted schemes (methods of assigning weights to events) and weights optimization (performance optimization by changing weight parameters). The paper shows that weighted methods don't lead to overfitting, i.e. learning on a training set doesn't decrease performance on an independent test set. We can see, that assemblers of different methods can increase performance (we consider linear combination of probabilities estimated by different methods). All experiments are made on real data of large International competition on data mining. The last span of statistics does not contain holidays, which allows concentrating only on statistical methods of problems solving while solving these tasks. Besides, we also considered construction of algorithm to solve the problems (next visit date and spends prediction) simultaneously. It can be seen that the problems not always can be solved independently. We propose a function to estimate solutions of both problems and optimization method for this function.

Citation: D’yakonov Alexander Gennad'evich (2014) Prognoz povedeniya klientov supermarketov s pomoshch'yu vesovykh skhem otsenok veroyatnostey i plotnostey [Supermarkets clients behaviour forecasting by weighted methods of probability and density estimations] Biznes-informatika, 1 (27), pp. 68-77 (in Russian)
BiBTeX
RIS
 
 
Rambler's Top100 rss