Studies that involve human factor modeling

---

"Technical" applications and research

---

 
Risk evaluation for a human car driver

NEXYAD analyzed how the human driver evaluates the risk (of potential accident), and developed algorithms that estimate the risk in real time (on-board application).

The driver can then be aware of the potential risk of accident.

Sensory analysis (touch, smell, sight, taste, ...)

NEXYAD worked on many sensory analysis projects : touch, smell, taste, acceleration feeling, ... For instance, NEXYAD worked on touch feeling of steering wheels covered with various materials. NEXYAD also worked on SMELL studies for car industry : car cockpit air filters efficiency quoting (these filters are made to cancel outside odours). In this framework NEXYAD developed a special know-how on protocole of measurement and experimentation (design plans, ...).

Odour sensor

NEXYAD developed a prototype of an odour gas sensor : it integrates FIGARO (metal oxyde gas sensors), pressure sensors, and temperature sensors. The biggest work was to develop a forecasting methods that compounds the very long time constant of gas sensors (5 minutes for absorption, 30 minutes for complete desorption). The non linear response of gas sensors was also used in order to simulate several sensors from one physical sensor. Classification is made with a neural net.

"Emotion" sensor

NEXYAD developed a prototype of an "emotion" sensor : several conductivity sensors are used by a classifier that builds classes of "mind state". This was used for the classification of mind states of a professional sniper (before the shoot, while shooting, after the shoot).

Expert systems development

NEXYAD developed many expert systems (knowledge-based systems) : it consists in modeling the logic of qualitative reasonning of experts in their field of application, and then in developing a computer program that can reproduce this qualitative reasoning.

Experts systems are mostly developed for industrial projects (process monitoring and piloting for instance). NEXYAD also applied expert systems to marketing and sales projects where there are many qualitative items of knowledge.

Driver's vigilance monitoring

NEXYAD worked on three approaches :  - driver's head monitoring, through eye tracking and blink detection,  - driver's behaviour monitoring : ethologists observe in situ the driver and count every behaviour occurrence. In the same time, every piloting signals are recorded (stiring wheel movement, pedals, ...). The  idea  is

then to classify the signals (instrumented car) and build classes of driver exhaust.  - vigilance level monitoring using conductivity measurements inside the driver's hands.

Driver's "style" monitoring

NEXYAD worked on the classification of driver's style. This classification may be used in order to adapt presets of suspensions to the need of the customer (very comfortable when the driver is cruising, very sport if the driver is piloting, security presets if dynamic signals are

recognized as a pre crash situation). It may also be used to feed the marketing teams with "real measured data on customers behaviour" (that may be correlated to declaration).

---

Marketing applications and research

---

Customers segmentation

NEXYAD developed a segmentation method that filters the declarations of customers through the actually bough product characteristics (example : ponderation of a customer who would declare 'i don't mind the price" & "i want speed" that would have bought a cheap slow car). This method automatically ponderate the items that don't bring homogeousty of the needs. As a matter of fact, this method leads to very homogenous groups of customers (in terms of needs). Example of application : segmentation of the small car market in Europe (cf. publication 1997 with the firm PSA Peugeot Citroën)

Satisfaction survey analysis

NEXYAD developed 2 new methods for satisfaction survey analysis :  - estimation of weight for satisfaction criteria : it is a "classical" linear way of modeling the satisfaction of customers. The idea is to give a weight to every satisfaction  criteria (summ of every weights = 1). The problem seems to have an obvious maths solution : linear regression. Unfortunately, linear regression (even pls or ridge regressions) often lead to negative weights (that have no meaning : they are an artefact due to strong correlations between variables). NEXYAD recently (2004) definitively solved this problem with a formal new algorithm (no more negative weights, formal additivity property of weights, ...).  - non linear modeling : the above linear modeling presumes that there exist a "compensation" effect between satisfaction criteria : "if i am very glad of this, and very annoyed by that ... i will have an average level of global satisfaction". NEXYAD could show that linear models are usually effective for customers that have a low global rate of unsatisfaction.

But because they are those who could stop to buy, unsatisfacted customers are very important to take into account. NEXYAD worked on very efficient decision models : criteria of satisfaction are not all equivalent : some HAVE to be satisfied at a very high level while others can stand being partially satisfied. There is no compensation effect, except between some special criteria : the unsatisfaction for ONE criteria may lead to a complete unsatisfaction for the global product (and then to a decision of no buy anymore). The use of such a model brings high correlation (for instance : 95%) between the declared global satisfaction and the predicted global satisfaction (predicted from all declared satisfaction criteria) when the linear modeling is usually poor (example : 65%). Such a non linear model is also very easy to understand in a marketing point of view (what to change in order to increase the global satisfaction for a cost as low as possible ? ...).

Attrition detection

NEXYAD developed an efficient method of sequence coding that was applied with succes to attrition detection (how to understand that a customer is about to move for a competitor ?).

Customers that are about to quit are detected, and those who are considered as being "interesting" are contacted for a "special offer".

Questionnary denoising

Studies lead to data files that are usually very noisy. This noise in not easy to model (it is not the result of a physical interaction between a sensor and an experiment) because there are several psychological effects :  - taboo : direct questions about some items may lead to no answers (example, in France : how much do you earn ?). Normally this effect is not too big in surveys because psychologists perfectly know this effect, and theyr avoid to ask direct questions.  - halation effect : customers don't forget their answer from a question

to the next question. They keep in mind what they answered and modulate their current answer with their previous answers. For instance, if a customer is very angry about something, it will lead to bad marks, but not only concerning the right item : every questions asked just after the item will be answered with angryness ... Usually surveys are printed in 4 or 5 versions (rank of questions is modified for every version). From this multi-versions questionnary, NEXYAD can detect halation and cancel this noise. Of course, statistics made from denoised data are much more relevant than statistics made on rough data.

Multi-country file fusion

Average levels for a given questions (example, a question like "do you ...", with answers like "never, sometimes, yes, always") may be different in different countries for several reasons :  - because the people of a country don't have the same use of the product, don't have the same needs, don't have the same values,  - because the people of a country don't understand the meaning of a question (their culture make the question irrelevant, forbidden, ...),  - because the use for answering is different  in a country (a country should answer between 5 and 10 because it is not polite to give a mark below 5 ...).

All these differences of culture make the file fusion a very complex problem : only the differences due to difference of needs should be kept, the other elements should be re-ajusted. Another complex variable is the PRICE of the bought product : levels of price are different from a country to another (food, real estate, ...) and implication into the studied product may also be different from a country to another (ex : in a country , people use to spend more than 50% of their incomes for food ? ...). NEXYAD developed methods and tools that help for data fusion.