Recommender systems for the modern digital workplace

Enterprise data is mostly unstructured and follows a siloed logic, which leaves little opportunity for meaningful metadata to occur. As a result, what would be an otherwise treasure trove of data, can hardly be used for automation or intelligent applications, nor can it be networked with more highly structured data in the enterprise. On top of all these headaches, search applications in most enterprises are still far more rudimentary than what we are used to from internet search engines like Google.

A recommender system is one such engine equipped to better support enterprises on their way to findability of assets and actionable insights. Positioned in the domain of knowledge-based recommenders, an enterprise recommender system that is built on Semantic AI has the potential to completely transform the way employees can interact with information. 

A key distinction between semantic recommender systems and the B2C recommenders that consumers are used to (think Amazon or Netflix) is that they do not produce trivial ‘more of the same’ information, but understand to some extent how information actually fits together.

How Semantic AI provides “secure” recommendations

Semantic AI has always focused on being able to infer and classify the meaning of large content sets in a largely automated way. Only when a machine is able to reliably annotate and disambiguate the meaning of a text in an explainable way, can it be used for further automation and applications, including in business-critical processes and decision making.

Currently on the market we see a desire for large language models (LLMs) that can answer questions and provide content based on the user’s prompt. ChatGPT is the first that comes to mind, but with many of its own drawbacks. Here is the fundamental problem with ChatGPT: it can provide answers and information that no one ever knows for sure is true because it is not referenceable. The sources and genesis behind the information always remains hidden. ChatGPT cannot remember how it knows something and does not know why it gives its answers.

There is no factual fidelity and so the clear danger is that users cannot rely on this information to make decisions. Especially as reputational risk becomes more at the forefront of enterprise success, companies bear a responsibility to be honest and transparent with their operations – and relying on systems that are neither traceable nor explainable indeed presents that risk.

Semantic AI is the remedy to this: it helps complement black-box machine learning or LLM with explainability and takes recommendations to the next level.

The core tenet of a semantic recommender system is that it is built on a company’s own knowledge model. This model is built by the subject matter experts and knowledge workers of the company who know their own data and can capture company knowledge in a domain that represents their company landscape, language, and more. 

Therefore, the recommender system is built on data they can trust and the recommendations that are spit out are more precise because they are generated from actual company data and content, and not from a third party source. In other words, the system is tailor made for the enterprise.

Increasing the quality of recommendations 

How does one distinguish if a recommendation is actually useful or not? A rather common experience is being on an e-commerce site and getting recommendations for a product that was just purchased; e.g. someone purchases an SD card for a camera and the website recommends another SD card. This person likely does not need two SD cards but could be enticed to buy a tripod, for example, an extra battery pack, a new camera lens. Each of these products are an opportunity for an upsell, and the less sophisticated recommender system is too busy suggesting things that have already been bought.

Consider this in your own company. If the search engine your customer website uses is operating on a similar level, do you really want to risk missing out on potential sales simply because the recommendations aren’t intelligent enough? 

A semantic recommender system does not offer more of the same, rather it provides sophisticated pairings based on the metadata (semantic footprints) that are defined in the knowledge model and related in the knowledge graph.

Intelligent recommender systems based on semantic footprints 

Semantic footprints embedded in domain semantic knowledge models can be used for accurate recommender systems that are superior to traditional vector-based similarity algorithms, since distances between data objects are not computed based only on the occurrence of terms.

The following example is taken from the demo application HR-Recommender, which is freely available on the web, and shows how relevant positions for employees can be calculated based on their weighted skills profile. A semantic knowledge model in the background helps to establish connections between occupations, skills and competencies that were not apparent at first, thus increasing the precision of matchmaking.

Consequently, in addition to recommendation systems based on “similarity,” PoolParty supports the creation of recommendation systems built on matching rules. Such rules are part of the semantic knowledge model and help to identify objects of different types that match or even complement each other. In this way, even challenging ‘configuration problems’ can be solved, e.g., to support cross-selling on e-commerce platforms. 

This example shows how the PoolParty recommendation system can be used to find competence profiles that are a good match for open positions and projects. Only employee biographies and job or project descriptions are required as input. The relevant skills, competencies and professions are automatically extracted from these documents and compared based on a knowledge model to automatically identify good matching pairs.

Internally, recommender systems help employees assemble information about a subject or project by retrieving the content they seek plus additional helpful information via these content pairings. Employees can better complete tasks and make decisions since they have a more comprehensive view of the matter at hand.

The PoolParty Recommender

Finding documents or other data objects similar or related to a given one is a highly desired feature to support content authoring, information retrieval, or configuration tasks. PoolParty provides various similarity algorithms and services that exploit semantic knowledge graphs and/or extracted features of a document. As an example, users benefit from a ‘Show similar documents’ functionality while searching for relevant documents. PoolParty provides semantic recommender services to be included in any analytics or search application. 

An upcoming release of PoolParty will include the unveiling of the Recommender Workbench which allows users to easily configure the output of PoolParty Recommenders. CEO Andreas Blumauer and Data & Knowledge Engineer Lorenzo Tosi will speak about this plus our recommender approach in a webinar being held on 17 October, where viewers can see for themselves the benefits of a semantic recommender system.