Machine Learning Algorithms For Behavior Detection
Today, with more and more data available to help us gain insight into customer behavior, organizations are beginning to use data more intelligently and are able to improve customer experience. However, predicting customer behavior remains a serious challenge.
In this article, Coding compiler will try to explain how Machine Learning processes help to identify customer behavior patterns and better understand what the client wants to receive or what his next step is.
Today, companies are trying to solve many issues and find solutions to business problems by smart data analysis but still limited in what they can do with their data, so consider the next step.
Organizations often analyze data using traditional information systems and maximize their data warehouse, but are still looking for a way to generate more and be able to use data in innovative ways. Today the goal in most cases is to predict the next step of the customer and thus add real business value.
There is a lot of confusion in what actually Data Scientist do?
A data scientist is someone who can make a connection between what happened “before” and what will happen “after”, and that includes:
- Identify new data sources, because traditional data analysis or traditional databases are unable to access them due to different format, size, or structure.
- Collect, match, and analyze data from multiple data sources.
- Application of the correct algorithm to get value from the data.
Ultimately, the goal of the Customer Behavior Identification Revolution is to:
- Find more information about the customer
- Implementing new processes to highlight the company expertise
- Use a new learning process to change your business
To illustrate this complex process , consider a mobile credit card application , which displays advertising ads related to customer purchases through the card-to-link platform, which is the platform that displays advertising ads and decides which ad is relevant to a specific customer.
Advertisers are pharmacies, restaurants, shops, etc., and they want to show ads to potential customers.
In order to target ads to relevant customers, you must accurately predict the likelihood that the customer will click on the ad – a click-through rate or CTR, also known as a clickthrough rate.
Card-to-link platforms are able to show purchase-related ads and will still strive to do a better job of targeting and measuring the success of their ads. On the other hand, the advertiser also wants to find new customers.
In addition to our example – for the advertiser – pizzeria , card-to-link platforms will connect the product to the customer through consumer behavior analysis and show the advertisement to a relevant customer, it will even find the customers closest to the pizzeria (based on the address of the existing customer’s address). And it works great, but if you set a more complex goal: the existing system increased new customers by 10%, but the pizzeria wants to increase its new customers by 30%. What should be done in this case?
Only through traditional data analysis cannot provide the desired results. Customer behavior needs to be predicted in a new way from the use of new/additional data sources to the running of various Machine Learning models. Let us try to elaborate.
The first question of Data Scientist is what type of data does the card-to-link have and can be used to increase the number of new customers from 10% to 30% thereby maximizing the profit of this campaign?
The first step is obtaining information. The new data comes from web browsing history; And are comprised of billions of ID devices and keywords in the browsing history for each ID. So how can a Data Scientist take the new data and locate a larger audience?
Data transfer. Since the data volumes are enormous, the platform faces the Big Data challenge, and here is a solution for NoSQL technologies such as Hadoop. In fact, all information about customer purchases, history of browsing, and history of the campaign is exported from a traditional data warehouse to NoSQL platforms. The collected data is used for interactive investigation and pre-processing of data using the SQL query engine.
Feature Extraction must now be performed – finding unique features in the data that can be used to make predictions.
Feature Engineering is a process of transforming raw data into a usable Machine Learning algorithm. The current tool for running models in Big Data environment is Spark using Spark MLlib. To use Spark algorithms, attributes must be arranged into Feature Vectors – number vectors that represent value for each attribute. To build a classifier model, you must run extraction and testing processes to locate the interest attributes that contribute to the classification. Data is also the text that can be processed, and it can also be converted into vectors of numbers.
The system can identify the most important words within the document and compare all documents. This is done by checking the number of times a word appears in a particular document and several times that the word appears in a series or collection of documents. For example, if there was a collection of documents on football, then the word “concussion” in the document would be more relevant for the document than the word “football”.
Running a Machine Learning Model.
For example, we are talking about a classification model . This is a family of algorithms that identify the category to which an item belongs (such as whether a customer likes pizza or not), based on marked data (such as purchase history). The classification model takes a set of tagged data and attributes and learns how to tag new records based on this information.
In our example, purchase history is used to tag customers who have purchased pizza, and the browsing history with millions of keywords that usually have nothing to do with pizza is used to create features that display similarities and categorize by customer types.
Once the browsing and acquisition histories are categorized as vectors, an algorithm such as regression or decision tree or random forests will be used to construct an appropriate model that “knows / learn” about the existing data and can predict the new data. The model can be used to rank ID’s of high-to-small pizza lovers. This result will give us the potential pizza lovers and thus can achieve a goal of growing customers.
Step Five – The Stage of After
Following the presentation of a model, analysis of data and forecasting, a campaign will be decided upon in advertisements for different audiences according to their consumer behavior. The campaigns will be presented to different target populations and it will be possible to compare the results of the campaigns with the forecast provided in the running of the model.
In our example, we will prepare a graph of campaign results and categorize customers according to their purchasing behavior, of course, according to the classification of populations created by the algorithm. For example, these will be customers who:
- They purchased pizza when they arrived at the pizzeria
- Buy pizza at a restaurant
- Purchase pizza in delivery
- Buy frozen pizza at the supermarket
The classified customers will be described using lines of different colors.
- On the Y axis (the vertical axis) we’ll show a percentage of customers who clicked on an ad in the app.
- On the X axis (the horizontal axis) we will show the probability that a person will click on the ad.
If color lines continue to rise from left to right, this means that the model worked well when predicting clickthrough rates (CTRs). If a particular customer type appears to have a decrease in clickthrough rates compared with the high probability, it can be concluded that the campaign failed. This shows that browsing history does not actually include increased clickthrough rates (CTRs) for the specific campaign, and the click-to-link platform could not identify it.
These are insights you can use, for example, if a click-to-link has a different campaign model or a new / additional campaign campaign decision. This is a method that works well and can be used for advertisers who pay to target ads to more customers.
In this article, we discussed how Data Science can use customers’ behavioral data, optimize advertising campaigns and bring more customers to the company. In other articles, we will talk about other uses of Machine Learning. Happy learning.
Related Data science Articles: