Neolink: Lifelong pathway recommendation challenges

Neolink is a company created in 2012 which developed solution to encourage the internet that brings together. Its main objective is to unify economic activity and social utility in its solutions. The company has been bought by Berger-Levrault in 2019.

Social services are notoriously complex to model and manage. It is essential to turn to mutualized systems that are tailored for the actors of these services. While examples of tools in the general public are legion today, there are still very few dedicated solutions in the social work and health sector, and even fewer offering long-term care for users. The objective of this project is to propose a multi-stakeholder and interactive recommendation system allowing to recommend social projects in the form of ordered set of actions, which we qualify as lifelong pathway.
This first part article is dedicated to the challenges in lifelong pathways recommendation.

Let’s consider the example of Alice as pictured in Figure 1 who wants to know how to achieve her personal long-term goal (apply to an accountant position).

Alice has already achieved a diagnosis of her current situation, similar to previous beneficiaries for which diagnoses and pathways were observed beforehand. The objective of the Pathways Recommender System (PRS) is to determine from there what would be the most relevant next actions for Alice to undertake to
achieve her long-term goal. This is a complex problem as it is heavily constrained:

• Alice has a limited time budget, and has to follow a certain learning curve so that each new action can easily build over the previously achieved actions.
• As a consequence, PRS has to estimate from previously observed beneficiaries, diagnoses, actions and their respective pathways and what would be the cost and the relevance for Alice to undertake an action.
• Similarly, the PRS needs to estimate a distance between actions to smooth the learning curve of Alice. Finally, the number and the diversity of actions make this problem difficult to solve.

Challenges in lifelong pathways recommendation

We consider in this work the problem of building a recommendation system to support social actors and beneficiary users in the interactive co-construction of a personal lifelong project. To illustrate it, we used the French active solidarity income (RSA).
These beneficiaries are supervised by social actors in order to build their project to improve their social integration or their financial situation. Until now, these procedures and this support are carried out mostly manually and administered by each administrative region. As these procedures are complex, regions turn to alternatives allowing them to facilitate certain aspects of this system. In this context, the Neolink company proposes a tool dedicated to the social actors to support the beneficiaries of this social benefit in the definition of their socio-professional project including the search for a job. This multi-stakeholders system is currently designed to allow the different social actors to collaborate and continue their work in a centralized manner. As such, the objectives are to enrich user experience and stakeholders interaction by mutualizing the best practices among the social actors and by helping them build a personal project for the beneficiaries from a very large set of possible actions related to health, social or professional aspects.

Research challenges

Several research challenges were expected for devising such a complex recommendation system:

1. Modeling the context of recommendation: what are the actors, their particularities and their interactions with one another and how are they expected to interact with the recommender system?
2. Dealing with the multi-stakeholders context: how to conciliate the objectives or constraints of each stakeholder among beneficiaries and the different social actors or services providers?
3. Formalizing the recommender output: we consider that lifelong pathways are specific composite actions related to health, professional or personal objectives of a beneficiary.
4. Providing a meaningful recommendation by determining the most adapted semantic and time granularity level for each action. Indeed, the expression of each action should be a balance between generality, for the ease of reuse, and the precision, to achieve meaningful recommendations for a specific beneficiary.
5. Ensuring the fairness of the recommendations and providing an explanation mechanism of it to increase trust of stakeholders in the proposition and to improve the acceptance of our solution.
6. Evaluating a multi-stakeholders, interactive lifelong pathways recommendation system, by determining the quality of the projects built by the system but also the quality of each recommended stage.

Recommender in its context

Figure 2 introduces an activity diagram of a model for lifelong pathways recommendation in the specific context of the RSA use case. This diagram composed of several steps, represents the processes observed from a careful study of our RSA data:

1. Initialization: a beneficiary enters the system and defines mid to long term objectives with the help of an expert that conducts a personal diagnosis, to start his lifelong pathways.
2. Stage: a supervisor co-constructs with the beneficiary a combination of actions to performed that is denoted as stage.
3. Report: actions taken during the previous stages are evaluated and the beneficiary path is updated accordingly. If some of the long-term objectives are not met, the whole process is reproduced from the diagnosis step in the Initialization layer.
4. Path end: Path closing, creation or a report about it.

Our main scientific challenges lie in the layer denoted Stage where a supervisor has to choose from a very large set of actions the ones to propose to the beneficiary. This step is the one for which we introduce the recommendation system that helps building stages. This recommender system has to take into account several stakeholders: the beneficiary, the supervisor and the expert as well as in the case of RSA, the funding entity and the preferences of action providers.
This leads to the introduction in our model of a new Recommendation layer for the automatic proposal of several combinations of actions to the supervisor. If these combinations match the context hard constraints (related to the date, the availability or the cost of the actions), the supervisor can establish a tailored proposal for the beneficiary based on those recommendations. As a consequence, recommendations have to be explainable since the supervisor may have to explain to the beneficiary how they were devised. Another challenge pertains to the interactions of the supervisor with the recommended stages, when selecting the most appropriate actions combinations, and how to transfer back this knowledge to the recommender system so as to propose more personalized recommendations in the future.
It is important to notice from Figure 2 that the objective of this research is not to replace any of the current social actors involved in the process but to facilitate their work in building better proposals with the beneficiary in mind. Final decisions are left to the supervisor and the beneficiary.

To be continued…