Introduction

Value-aware recommenders are systems that aim to directly maximise the economic value (e.g revenue, profit…) of recommendations by considering the objectives of multiple stakeholders. The typical applicaitions are online marketplace and sponsored Ad.

In this blog, let’s talk about some high-level overview. In the next blog, we’ll check more details about value-aware recommender, and specifically focus on online marketplace and the trade-off between user satisfaction and platform profitability.

MORS, MSRS & VARS: Definition and Taxonomy

Relationships between multi-objective recommender, multi-stakeholder recommender and value-aware recommender

Multi-Objective Recommender System (MORS)

Tradionally, recommenders are mainly optimized for single objective: user experience, namely item relevance. However, many platforms face multiple objectives, and joinly optimizing them remains a challenging task.

There are several ways to classify MORS, and it’s classfied to three main categories here:

Recommender with multiple quality metrics, where only objectives of end users are considered, but except relevance, other quality objective for users, such as novelty, diversity and serendipity, must be considerd.

The underlying assumption is that improving different quality aspects for users would be at least indirectly beneficial for the recommender service provider.

Multi-stakeholder recommender, where, except objective of end users, the objectives of other stakholder such as recommender service provider and item providers are considered. And the related objectives may stand in conflict

A typical problem setting involves two stakeholders: end users and recommender service provider. There is a potential trade-off between

Recommend the most relevant items for the end users
Recommend the item that are also somewhat relevant but assumed to be favorable in terms of the service provider’s business objectives

Multi-stakeholder recommender can enhance the benefits of other stakeholders without or with a small loss for the end users who are the receivers of the recommendations.

Recommender with different time horizon objective, where short-term and long-term objectives are both included

Example of long-term objectives are customer retention or customer life time value

Multi-Stakeholder Recommender System (MSRS)

In multi-stakeholder recommender system, there could be at least one objective associated with each stakeholder. MSRS is expected to achieve a balance among these stakeholders.

Depending on the type of business, MSRS can be classified into three categories:

Value-Aware Recommender, which aims to maximize the economic value of recommender. Examples are recommender in online marketplace and sponsored recommendation
Reciprocal recommendation, which recommends user to user instead of recomemending item to user. Typical application scenarios are recommendations on dating and recruiting platform.

The sucess of recommender is not determined solely by the recipient, but there must be a mutual preference match or compatibility between two people involved.

Group recommendation, where the goal is to recommend items for a group of users, for example, a group of friends who want to watch movie together

Value-Aware Recommender System (VARS)

For traditional recommender which only considers item relevance, the underlying assumption is that recommending relevant items to users can also have a positive impact on the value for the providers or the platform. In this way, economic value is influenced indirectly by recommender

However, a more direct way is adopted in VARS by considering economic value in recommender. And the assumption of VARS is: platform profitablity can be improved without a significant loss in terms of user satisfaction.

Typical application scenarios of VARS are

Online marketplace, where most of the time, objectives of three groups are included:

End consumers (customers/users with a need to be met by the marketplace and who receive recommendations to fulfill it)
Item providers (entities that provide goods/services to the customer through the marketplace)
Recommender service provider (the platform that matches consumers to providers).

Sponsored recommendation, where user preferences and Ad revenue should be both considered in the optimization process

Multi-Objective Optimization Problem (MOOP)

To solve the problem of multi-objective recommender, muliti-objective optimization is needed. Let’s get some basic idea about MOOP. In the next blog, we’ll talk about how MOOP is applied in value-aware recommender.

Problem Definition

Let’s define $F(x)$ as the vector function in objective space $R^M$ , MOOP can be expressed as

\min_x F(x) = \min_x (f_1(x),f_2(x),f_3(x)...f_M(x))

Subject to $g_k(x)\leq0, k=0, 1, 2...K$

Where

M : number of objective functions
$x=(x_1, x_2...x_n)$ : n-dimension decision vector in space $R^n$
$g_k(x)$ : possible constraints in MOOP

A feasible solution for the above problem is called Pareto optimal solution , which is a state when it’s not possible to improve one objective without hurting other objectives.

Optimization Methods

There are different ways to classify Multi-Objective Optimization methods. Based on underlying optimization strategy, they are classifed to two main categories:

Scalarization Methods:

Basic idea: transfom a multi-objective problem to a single-objective problem so that the existing solutions for single-objective problem can be used.

Heuristic Search:

Basic idea: optimize multiple objectives directly using mutl-objective evolutionary algorithm such as evolutionary algorithms or swarm intelligence method

While there are many researchs focusing on to apply heuristic search method to recommender. Due to its simplicity, scalarization method is used more in real-world recommender, and we’ll focus on this method in the next section.

Scalarization Methods

There are many ways to transform multi-objective problem to simple-objective problem, here we’ll discuss two most popular methods in multi-objective recommender system.

Weighted Sum Method

This method scalarize a set of objectives into a single objective by adding each objective premultiplied by a weight, which is straightforward and popular.

In this case, the objective is changed to

\min_x F(x) = \min_x(f_1(x),f_2(x),f_3(x)...f_M(x)) \\ = \min_x \sum_{i=1}^{M}w_if_i(x)

Where $w_i$ is the weight for each objective, and the weights are chosen in proportion to the relative importance of the objective. There are different variants for this method, such as

Weighted exponential sum
Weighted product
Weighted metric method
…..

In addition, despite the single-objective formulation, it is not guaranteed that the solution to the problem is Pareto optimal solution, unless proper weights are assigned.

$\epsilon$ -Constraint Method

The second method is $\epsilon$ -constraint method, where it optimizes one objective and treats other objectvies as constraints:

\min_x F(x) = \min_x (f_1(x),f_2(x),f_3(x)...f_M(x)) \\ = \min_x f_l(x)

subject to: $f_i(x)\leq \epsilon_i$ , for all $i=1, 2...K, and \space i\neq l$

‣

Value-Aware Recommender System-Part 1: High Level Overview