Recommender System based on Predictiveworks. (Update)

December 7, 2014

Our recommender system based on Predictiveworks. is now ready for use.

Background: Recommenders have a strong focus on user item interactions, and can be derived by creating explicit user and item profiles to characterize their nature. These profiles allow to associate users with matching items and, from that data, specify user item interactions (content filtering).

An alternative approach that avoid the extensive gathering of external profile data is based on the extraction of inherent and latent data from the underlying dataset to specify users, items and their interactions. Our recommender uses latent data and supports two different approaches to personalized recommendation:

Latent Relations

This approach uncovers latent relations between items in a large scale dataset und combines these relations with the last user transaction data. It is based on Association Rule Mining and uses the Top-K NR algorithm from Philippe-Fournier Viger.

In 2012, Philippe-Fournier Viger redefined the problem of association rule mining as 'Top-K Association Rule Mining', and circumvents the well-known 'minimum support' problem.

Latent Factors

This approach learns latent factors for each user and item from the underlying dataset and uses these factors to specify user item interactions (matrix factorization). In addition to learning latent factors for users and items, the recommender is also capable uncovers factors for additional features that describe the context of a user item engagement.

The latter approach is based on factorization models. These models are a generalization of matrix factorization and are capable of modeling complex relationships in the data to provide context-aware personalized recommendations.

Readers that are interested in technical details may continue reading here.

Recommender System based on Predictiveworks.

November 28, 2014

We are proud to announce that Dr. Krusche & Partner has started a next-generation recommender system. It uses the stack of Predictiveworks. and combines different engines to address e.g. article and product recommendations with multiple machine learning algorithms. Among the feature highlights is implicit user preference computing and context-aware recommendations.

We keep you informed about this open source project in this blog.

Readers that are interested in technical details may continue reading here.

Shopify meets Predictiveworks.

November 18, 2014

Shopify provides an amazing software platform for both online stores and retail point-of-sale systems. The company helps emerging small business get off the ground and grow into successful companies.

We share this mission and have made Predictiveworks. to help small and mid-size companies boosting their business by leveraging cutting-edge data mining and predictive analytics capabilities.

Shopifyinsight bridges both worlds and makes it easy for shopify customers to apply Market Basket Analysis to their orders, compute Product Recommendations and more. Shopifyinsight uses Shopify's API to collect shopping data and provides access to on demand mining and prediction capabilities through its REST API.


Association Rules

November 16, 2014

Association rules are IF-THEN statements that uncover relationships between seemingly unrelated data in almost any data repository. An example of such a rule would be "If a customer buys a dozen eggs, he is 80% likely to also purchase milk."

An association rule has an antecedent (IF) and a consequent (THEN) part. An antecedent is an item(set) found in the data. A consequent is an item(set) that is found in combination with the antecedent.

Association rules are created by analyzing data for frequent IF-THEN patterns and using the criteria support and confidence to identify the most important relationships. Support is an indication of how frequently the items appear in the database. Confidence indicates the number of times the IF-THEN statements have been found to be true.

Association Analysis is useful for analyzing and predicting customer behavior. It plays an important part in shopping basket analysis, product clustering, catalog design and store layout. And, it may also be used for product recommendations.

For more information: Association Analysis

Series Analysis

November 11, 2014

Series Analysis brings the temporal dimension to customer behavior analytics. To know which products are bought one after another in a cycle helps to forecast product demand and sales, enables just in time recommendations, predicts which products will be bought next and when, and more.

E-commerce is a prominent use case for Series Analysis, but it is not the only one. Time-ordered sequences of activities and events occur in many other domains, such as finance, government, health care, to name just a few.

For more information: Series Analysis

Market Basket Analysis

November 6, 2014

Market Basket Analysis is a wide-spread mechanism used by retailers to gain a better understanding of customers’ purchase behavior. It improves cross- and up-selling, product placement, sales promotions, loyalty programs, store design, and discount plans.

Market Basket Analysis can also be applied for customer segmentation or divide products into natural categories.

A major difficulty of today's algorithms is that a large number of product rules is found that may be trivial for anyone familiar with the business. Requiring rules to have a high minimum support level and a high confidence level risks missing any exploitable result we might have found.

The Association Analysis engine of Predictiveworks uses an alternative approach and discovers the top most relevant rules and thus overcomes this well-known threshold problem. This leads to better decisions throughout the entire company, including:

  • Intelligent assortment decisions
  • Enhanced inventory readiness for promotions
  • Improved inventory distribution by store
  • Better item pairings in marketing materials and store displays
  • Enablement of insight-driven guided selling for clienteling and personal shopping
  • Complete analysis of promotion effectiveness

For more information: Association Analysis

Context-Aware Recommendations

November 3, 2014

Behavior is strongly influenced by the current setting and mood of a person. Information about the situation in which a choice or decision has been made is important to improve customer understanding and thus establish more personalized relationships.

Most of today's recommender systems rely on how (rating) who (user) rated what (article, product etc.). The situation (context) in which a user rating is made is out of scope. Today's recommendations thus reach a very limited degree of personalization. But, this is what recommender systems are usually made for.

Context-aware recommender system have a strong focus on the situation in which a choice is made and outperform existing recommenders. However, the few existing approaches suffer from severe drawbacks such as high model complexity, limited variable usage and low performance.


The Context-Aware Analysis engine of Predictiveworks uses a alternative approach: Factorization Machines. They have linear complexity, do not put any restrictions on the variable to use and provide a better prediction quality in much less time.

Factorization machines are not restricted to recommender systems and provide a general purpose approach to most context-aware problem by just feature engineering.

For more information: Context-Aware Analysis

Predictiveworks.  brings the power of predictions to Elasticsearch.

Association Analysis

Discover and leverage the most relevant
hidden relations in large scale datasets.

Context-Aware Analysis

Leveraging context-sensitive information
is crucial for personalized predictions.

Decision Analysis

Predict the best decisions among
multiple courses of action.

Intent Recognition

Uncover the intents of human behavior and reach the ultimate customer understanding.

Outlier Detection

Find anomalies in large-scale datasets and human behavior for advanced risk reduction.

Series Analysis

Discover and leverage the most relevant
patterns in human activity sequences.

Similarity Analysis

Find relevant similarities in activity sequences and identify customers by their journey.

Social Analysis

Determine and leverage actual trends
from social media in real-time.

Text Analysis

Language-agnostic semantic concept
extraction for semantic targeting.

Contact us   +49 (0) 89 898 27733   •   team@dr-kruscheundpartner.de