Authors: Rocío B. Ayala Meza
Climate change and geopolitical factors threaten the availability of essential to life resources such as water or energy in emerging and developed countries. By predicting risk of asset damage in urban (constructions, and commodities warehouses) and rural areas (natural and rural crops), it would be possible to make informed decisions and adjust essential to life needs.
Data-driven techniques, like Machine Learning (ML), have gained prominence in multiple applications thanks to the surging amount of data and computational power. Nevertheless, risk assessment is a complex problem, as it involves multiple criteria and sometimes this data is difficult to collect. Therefore, it is reasonable to adopt a Multiple Criteria Decision Making (MCDM) method like the Analytical Hierarchy Process (AHP) to tackle it.
Finz, the project organizers, focuses on helping global decision-makers to finance autonomous and light units for water or energy efficient access in rural and urban contexts. They entrusted this project to Omdena.
The delivered project consisted in the identification of flood and violence susceptible areas, in West Africa (Togo and Ivory Coast, respectively). Therefore, the purpose of this article is:
- To explain risk and the importance of assessing it.
- To present the AHP method and describe how it was used to solve this challenge.
To elaborate on how AHP and other MCDM methods can be combined with ML and why the knowledge of human experts (or stakeholders), sometimes, cannot be overlooked.
What risk? Why is it important to assess it?
Risk is the probability of a hazard occurrence. It is a combination of three components: hazard, exposure and vulnerability (See Image 1).
It can be presented conceptually with the following equation:
- Hazard: Describes the effect and probability of a phenomenon. Ex: floods, war.
- Probability: Chance of something happening.
- Effect: It is the consequence of an event.
- Vulnerability: It is the degree of susceptibility between the exposure and the hazard. Ex: buildings close to the seashore may be more vulnerable to the impacts of flooding.
- Exposure: These are the assets and population that can be affected by a hazard. Ex: buildings, farmland.
- Susceptibility: It is the lack of ability to resist some external event.
Risk information helps to know what, when and where a dangerous phenomenon might happen, how severe it could be, and who would be the most affected.
What is Analytic Hierarchy Process (AHP) and how does it work?
MCDM is a sub-discipline of Operations Research that evaluates multiple contradictory criteria in decision making. An example of conflicting criteria is the management of an investment portfolio. Because high gains are sought while minimizing risk, despite the fact that most profitable stocks are the riskier ones. For a list of other MCDM methods and how they are used in sustainable projects, see Table 2.
Analytic Hierarchy Process (AHP) is a MCDM method (proposed by Professor Thomas L. Saaty), based on mathematics and psychology, to make rational decisions by quantifying subjective beliefs. It can be applied in complex situations like quality management, conflict resolution and bench marking; and in diverse fields ranging from sustainable systems to portfolio selection.
This highly intuitive process consists of the following steps (See Image 2):
- STEP 1: Define the goal of the process, the various alternatives to reach that goal, and list the criteria to evaluate those choices.
- STEP 2: Build a hierarchy based on pairwise comparisons by imposing judgments. For every pair, assign a score from the Saaty scale (see Table1) to the most important option and the reciprocal value to the other one. Ex. Criteria A is 3 times more important than Criteria B, then we assign 3 and 1/3, respectively.
- STEP 3: Calculate the priority vector (Vp) for each criteria.
- Make a sum by column and divide each element, of the same column, by that sum.
Ex. In column A, Sum = 1 + 0.33 + 0.14 + 0.11 = 1.59.
Then Column A values will become 1/1.59, 0.33/1.59, 0.14/1.59, 0.11/1.59
- Make a sum by row and calculate the average by row
Ex. In row A, 1/4 * (1/1.59 + 3/4.34 + 7/13.33 + 9/20) = 0.5738
- Make a sum by column and divide each element, of the same column, by that sum.
- STEP 4: Verify the consistency of the judgments. A Consistency Ratio (CR) greater than 0.1, indicates the probability that the judgments were randomly generated, and therefore the AHP matrix must be revised.
- Calculate the Principal Eigenvalue (λmax). This is the summation of products between the sum of columns and each Vp.
- Calculate Consistency Index (CI) and CR
n: is the number of criteria. Ex. In Level 1 there are 4 criteria.
The Random Consistency Index (RI) is obtained from Table 2.
- STEP 5: Compute the overall composite weight of each choice based on the weight of level 1 and level 2. The one with the highest value will be the preferable choice.
Steps 2 to 4 have to be repeated for all the matrices in level 2 with respect to each criteria.
How was Analytic Hierarchy Process used to assess risk?
The project consisted of two domains: Climate change and Geopolitical risks. Each one used the AHP method with different criteria and factors (choices).
a) Climate change domain
Floods have an impact on human society and ecosystems because they threaten property, pollute water resources and change the natural environment. Flood risk is related to environmental and human factors like increase of rainfall and rapid urban development, respectively.
The goal (Level 0) was to assess flood risk in the country of Togo. The criteria (Level 1) were hazard and vulnerability, which are required indexes to estimate risk. Finally, the factors (Level 2) were elements, according to human experts, that influence flooding. For example, Precipitation is the main natural hazard responsible for it. All these factors and their weights (or influence) can be changed according to the experts’ opinions.
Flood data is rare, since they do not happen every day. Therefore, the Analytic Hierarchy Process (AHP) method was used to create data that was later fed into a ML model, which in future work can be used to create a risk score in the surrounding areas without creating a new AHP model nor training another ML model. Thus, reducing the computational cost.
But the initial AHP model can also be used to assess risk in adjacent areas without using ML. As it can be expanded to include more, or other, expert layers depending on the territory and hazard, like flow accumulation and surface roughness for flash flooding, or burn rates and vegetation types for bush fires.
b) Geopolitical domain
Geopolitics is how economics and geography influence relations and politics between countries. For example, terrorist attacks in the Middle East could result in a spike in oil prices. Geopolitical risk is mainly related to human factors like riots, war, violence against civilians, and protests.
The goal was to assess the risk of diverse violence factors (riots, wars, etc) for the country of Ivory Coast. The criteria were the event’s type, location and time. Each event has an individual risk score, then these are aggregated to produce a final risk score for a specific location (according to its Latitude and Longitude coordinates).
Here, the applied AHP method is a customized version because geopolitical risk score is not driven by physical events but is more a social economic problem. Therefore, there are many interpretations to what can be defined as a geopolitical risk factor. Also, many of these are anthropological in nature, making them difficult, if not impossible to measure.
Thus, a risk score equation was created. It is observed that weights w2 and w3 ensure events that happened further away, in both a spatial and temporal manner, have less effect on the overall risk score. Much like in the previous case, these weights can be tuned to adapt to the current situation and provide a reliable score.
What are the advantages and disadvantages of combining AHP with ML?
ML relies strongly on big data and computational power. Many ML applications are limited by a lack of data; either because the collection is difficult (Climate change factors) or simply because it is old or has quality issues (Geopolitical factors). Nevertheless, its reliance on big data, when applied to rare events, might be addressed with the Internet of Things (IoT).
AHP is useful when people are working on complex problems (of long-term repercussions) which involve human perception and judgment, as it provides a framework for the decision support process. When combined with ML, it provides interpretability to the models. Another advantage of this blending is that humans can refine the weights of the factors of interest (ex. precipitation and event type), and therefore provide a credible risk score.
Analytic Hierarchy Process (AHP) can also be used to label data; and thus create a ground-truth dataset ready to be applied with ML algorithms. Nevertheless, AHP is susceptible to human expert judgment. Thus, different studies for the same problem will identify different sets of criteria and choices; and these will not have the same influence on the goal.
Other MCDM methods and their application in sustainable projects
There are many MCDM methods, besides AHP, that offer experts the opportunity to contribute with their knowledge. For example, Technique for the Order of Prioritization by Similarity to Ideal Solution (TOPSIS), Preference Ranking Organization METHod for Enrichment of Evaluations (PROMETHEE), Multicriteria Optimization and Compromise Solution (VIKOR) and Analytic Network Process (ANP).
Climate change and Geopolitical factors pose serious threats to the availability of essential-to-life resources, such as water and energy. Decision-makers need to act and decide with accuracy in predicting risk on assets or population.
In this project AHP, a MCDM method, was used to assess Climate change (floods) and Geopolitical risk (riots, wars, etc) in the countries of Togo and Ivory Coast respectively. For flood risk assessment, a hybrid method was developed (AHP-ML). The AHP method was used to create ground truth data that later served as an input to many ML models. Meanwhile, in order to quantify Geopolitical risk, an equation (inspired by the functioning of AHP) was formulated.
There are many MCDM methods that can be combined with ML models. Some advantages are that models will have interpretability and that stakeholders can refine results by providing their expert knowledge. Also, some types of events are rare, and therefore data is scarce. In those cases, MCDM methods combined with remote sensing can be used to create ground truth datasets for ML analysis.
Based on the obtained results, AHP and other MCDM methods can be used in future Omdena challenges.
Building AI Solutions for Real-World Problems