reference cases Data Science

MANY DIFFERENT APPLICATION AREAS

Data Science and its general methods can be successfully used in a very wide range of areas. Advectas works independently of the individual industry and operational area. Here are a few examples of areas where we can contribute our expertise.

sales & marketing

 

predict churn

challenge
Some customers are more likely than others to switch supplier (price hunters). The customer commits to a new supplier, which makes win-backs tough.

solution
Predict whether a customer will terminate its contract. The prediction is based on factors such as type of agreement, consumption patterns, and interactions.

Tech
Classification with a XGBoost model based on a snapshot of recent interactions with the company, type of agreement, consumption, etc.

business value
The company can now prioritize its marketing better for retaining the customer base, rather than winning back they who have already left. They have also made themselves more relevant to customers by providing customized offers at the right time.

FORECAST NUMBER OF VISITORS

challenge
The number of visitors is Liseberg’s most important KPI. It’s strongly dependent on external factors such as weather, other events in Gothenburg, season, weekday, holidays, etc. But also internal factors such as is there an artist playing on the big stage, is there a theme in the park, opening hours, etc.

solution
Predict the number of visitors per hour, 10 days ahead, based on historical internal data as well as external data sources (such as SMHI and last.fm).

Tech
Facebook Prophet was used for day-level predictions, and a neural network in PyTorch distributed the visitors by hour. The solution was put into production in a Docker container in Azure.

business value
The forecasts is a basis for marketing and personnel scheduling. With better planning, the Liseberg visitor will have a better experience.

PREDICT CROSS-SALES

challenge
Some customers only buy a few types of the bank’s products / services. Without knowing the customer’s preferences, it is difficult to be relevant in their marketing.

solution
Twin-based model based on customer data, historical transactions and activities. Consumption of different product groups per customer was predicted as a sales-potential and then compared with actual sales to find unutilized potential.

Tech
LightGBM model written in Python and deployed on a Microsoft SQL-server.

business value
Targeted campaigns and offers to customers with high purchasing potential per product / service instead of broad campaigns in order to become more relevant to customers with increased sales as a result.

MARKET BASKET ANALYSIS

challenge
Customers expect that offers and promotions are always relevant. If you have a larger assortment it will quickly become complex and difficult to find combinations of products that are usually sold together. You also want to know if a product drives sales of other products.

solution
Statistical calculations to find which products are bought together and different association rules; which products drive positive or negative sales of others. Ability to filter out sales during desired periods.

Tech
Developed in Python with the Dask framework for parallelizing data prep. To find combinations of products, bindings to C ++ and algorithms like Apriori and FPgrowth were used.

business value
Decision support for creating and following up campaigns. Better understanding of sales and complex patterns that can arise and be utilized.

PRICE OPTIMIzATION

challenge
The biggest factor for most customers within retail is the price of the products. It’s a challenge to quickly and with precision follow and analyze how price changes affect sales volumes and turnover.

solution
A Dashboard easy to use that enables category managers to do What-if analyzes to understand how much or how less they would sell depending on the price they set for a product.

Tech
A connection to the Data Warehouse to be able to read and analyze receipt rows. Price elasticity calculations is done in Python. Dashboard delivered in Tableau, Power BI or any other BI-tool.

business value
The company is now able to do What-if analyses before they set the price. The prognosis for a product will be distributed over the seasons. The company is also able to quickly follow-up and analyze the effects of the price changes they make. This solution can also be used for pricing your markdown.

PRISOPTIMERING

OPERATIONS

PREDICT REINSERTION OF PATIENTS

challenge
In this region, psychiatric care is under heavy strain. Several doctors need to agree on priorities between patients. Broken care chains (patients being discharged prematurely and relapsed) result in poorer quality, longer care time and higher costs. Large amounts of text in journals make it difficult to create an overview and insight into longer care cases and chains of care.

solution
Predict probability that a patient will return for more care in the near future based on, among other things, free text in medical records.

Tech
Create mathematical representation of journal texts with Doc2Vec models. Then Gradient Boosted Decision Trees (XGBoost) for classification.

business value
The doctors of psychiatry can use the prediction result as a second opinion on prioritization between patients.

HR & STAFFING

SUMMARY OF SURVEYS/REVIEWS

challenge
Receives several thousands of survey responses every year. Analyzing these are very time-consuming and human bias might affect the result. Often, only a small selection of texts is read to try to form a holistic view.

solution
Find recurring topics (groups of words that are often used together). This is a statistical approach which is independent of language. It’s also possible to track topic magnitude over time to follow up actions.

Tech
Topic modeling with LDA (Latent Dirichlet Allocation) and NMF (Non-negative matrix factorization) developed in Python. Visualized in several different BI tools as well as standalone HTML.

business value
Impartial summary of texts. Big time savings. Ability to identify and act on insights that otherwise drown in large masses of text.

OPTIMIZE SHIFT SCHEDULING

challenge
24/7 operations and other staff-intensive operations have a great need to put in place reliable and resource efficient shift schedules. With tools like Excel, the process is time consuming and the risk is great for mistakes and sub-optimality.

solution
With decision optimization, the process can be streamlined. Together with machine learning, demand can first be forecasted and the schedule optimized automatically thereafter.

Tech
Developed in Python with optimization engines like CPLEX.

business value
An automated scheduling process can save countless hours of administrative time. An optimal schedule also ensures an efficient use of resources.