Диссертация (1137066), страница 3
Текст из файла (страница 3)
Neurons are special cells that can propagate electrochemical signals. The neuron has a branched information input structure (dendrites), a core anda branched outlet (axon). Axons cells connect to the dendrites of other cells by using synapses. Upon activation the neuron sends an electrochemical signal along itsaxon. Through synapses this signal reaches other neurons, which can be activated inturn. The neuron is activated when the total level of signals, come to its core fromdendrites, exceeds a certain level (an activation threshold).It is interesting that artificial neural networks are able to achieve remarkableresults using a model that is not much more complicated than that described above[12,14].To reflect the essence of biological neural systems, we note its characteristics.The neuron receives input signals (initial data or the output signals of other neuronsof neural network) through several input channels.
Each input signal passes througha connection having a certain intensity (or weight); this weight corresponds to thesynaptic activity of the biological neuron. Each neuron is associated with a certainthreshold value. The weighted sum of the inputs is calculated, the threshold value issubtracted from it and, as a result, the activation value of the neuron is obtained (it isalso called the post-synaptic potential of the neuron-PSP).The activation signal is converted using the activation function (or transfer function) and as a result, the output signal of the neuron is obtained.If one uses the stepped activation function (i.e., if the input is negative, the outputof the neuron is equal to zero, and one if the input is zero or positive, the output isequal to one), then such neuron will work just like the natural neuron described above(to subtract the threshold value from the weighted sum and compare the result withzero is the same as comparing the weighted sum with the threshold value).
In fact,threshold functions are rarely used in artificial neural networks [18]. It is noted thatweights can be negative, it means that the synapse renders on neuron not the exciting,but braking influence (there are braking neurons in the brain).The network takes inputs (accepting values of variables from the outside world)and produces outputs (forecasts or the controlling signals). Inputs and outputs cor11respond to sensor and motive nerves, for example, respectively, going from eyes tohands. Besides, however, there may be a lot of intermediate (hidden) neurons in thenetwork that perform internal functions.
The input, hidden and output neurons areconnected among themselves.Key question here is backlink. The simplest network has a direct signal transmission structure: The signals pass from the inputs through the hidden elements andeventually arrive at the output elements. Such structure has steady behavior. If thenetwork is recurrent (i.e. it contains connections leading back from more distant tomore near neurons), then it can be unstable and have very complex dynamics of behavior.
At the same time recurrent networks are of greater interest to researchers inthe field of time series data.Any organization in its activities, seeks to maximize profit. Banks are no exception. To do this, they need to organize the work, focusing on those things thatbring the greatest profit with the lowest cost.
It was believed that the product or theorganization’s services is the main profit driver. However, recently client-orientedstrategy has gained popularity, where the client and the mechanisms of interactionwith him/her are the key to successful sales. This strategy is called CRM - CustomerRelationship Management [25].To attract new clients banks use both passive advertising, for example, advertising on television, radio, press and direct mailing commercial offers. To increasethe effectiveness of such measures it should take into account the interests of clients,impact of objects, i.e. to offer clients the product that they prefer. But it is impossibleto take into account the preferences of each client. It is necessary to allocate somegroups–segments of clients and to propose specific category of financial products andservices for these groups.It is possible to allocate client segments on several groups of signs.
These canbe segments by area of activity, by geographical location. After segmentation it ispossible to learn what segments are the most active, what make the greatest profit,where there are most loyal clients and to highlight their specific features. To solvethis task it is applied a powerful clustering mechanism - self-organizing maps ofKohonen.12One of the most likely reasons neural networks have not had many applicationsin credit scoring is that they cannot be interpreted.
At the same time some regulatorsrequire this property. For instance, in 28 October 1974 Equal Credit Opportunity Act(ECOA) was adopted in the USA. The act prohibits credit discrimination on the basisof race, religion, sex, marital status, age, or because you get public assistance. Creditors may ask borrowers for most of this information in certain situations, but theymay not use it when deciding whether to grant a loan or when setting the terms of aloan.
Also consumers are able to require the reject reason. Therefore, mathematicalmodel used in credit scoring in most cases has to be interpretable. Nevertheless, experiments with sophisticated machine learning algorithms in credit scoring reflect theresearch for opportunities to improve simple logistic regression models performancein this domain.The goal of this Ph.D. thesis is to achieve better performance but still beingable to interpret the model, and throughout following sections we will show howinterval pattern structures allow one to increase model predictive power still havinginterpretable rules for borrower classification.2.3Classification Tasks in Marketing Campaign ManagementPresently financial services companies have gained new opportunities thanks tomachine learning that represents a number of the processes allocating computers withability of making hypotheses based on the known characteristics taken from trainingdata.Even more often companies operating in the sphere of crediting use machinetraining for forecasting of solvency of clients and also for creation of models ofcredit risks.
Among such companies — Kabbage, Inc., financing small businessby platform of crediting, service of remote microcredit LendUp and the recognizedleader of branch of financial Lending Club technologies. In particular, the Kabbageteam specializes in development of algorithms of machine learning of new generation and analytics for creation of models of credit risk and the analysis of the existingportfolio. Among a set of machine learning algorithms for determination of rating13of solvency of the borrower the following is used: multilayer perceptron, logistic regression, Support Vector Machine and as well as AdaBoost (or Adaptive Boosting)classifiers.Financial computing and decision making can be done through machine learning algorithms that allow computers to process data more efficiently and more quicklyand make decisions about crediting, insurance, fraud protection, etc.
Machine learning models are widely used by companies such as Affirm, BillGuard and ZestFinance[5]. The latter has managed to find a new approach to traditional tasks thanks to machine learning and analysis of large data sets. Company analyzes thousands of potential credit variables - from financial information to the use of technology, to betterassess factors such as potential fraud, the risk of default and the probability of longterm relationships with clients. As a result, the enterprise can make more "correct"decisions of loan granting, which leads to increased credit availability for borrowersand higher percentage of their maturity.Cross-sell, up-sell and churn modeling are one of the traditional and well-developedapplied areas of machine learning in marketing. Effective solutions can be receivedwith use of decision trees, regression analysis and other analytical methods and models.Usually the integral indicator of each client is compared by a numerical threshold, which in essence is the level of profitability and is calculated from the ratio,how many on an average need clients who pay on time to compensate losses fromone debtor.