Traditional recommender systems operate on a simple premise: calculate a utility score for every individual item in a catalog, sort them in descending order, and serve the top results as a scrollable list. While this approach works well for independent choices—like picking a single movie to watch next—it breaks down when items are meant to be consumed together.
If a user is looking for an outfit, a traditional system might recommend five highly-rated shirts, ignoring the fact that the user actually needs a matching pair of pants, shoes, and a jacket. To solve this, a growing body of research shifts the focus from ranking individual, atomic items to generating and completing cohesive sets, packages, or bundles.
Beyond Independent Rankings: The Portfolio Effect
The core limitation of standard recommendation models is that they treat items as independent entities. As Marie desJardins et al. demonstrated in Learning user preferences for sets of objects, evaluating items in isolation fails to account for the portfolio effect—the idea that the value of a collection depends fundamentally on the interactions between its members, such as their complementarity or redundancy.
When users make complex decisions, they are often looking for a "package" that fulfills a broader, multi-stage goal. For example, Minzhu Xie et al. formalized this in Breaking out of the box of recommendations: from items to packages, noting that real-world tasks like trip planning require collections of items that collectively satisfy a user's needs while adhering to global constraints like time or cost.
Traditional Ranking (Item Isolation):
[Item A: 9.5] -> [Item B: 9.4] -> [Item C: 9.3] (e.g., 3 identical black shoes)
Set-Based Generation (Contextual Coherence):
{ Item A (Shoes), Item F (Pants), Item K (Shirt) } -> High Collective Utility
To build systems capable of generating these collections, researchers have had to redefine the relationship between items. As Yong-Siang Shih et al. argue in Compatibility Family Learning for Item Recommendation and Generation, the fundamental challenge is modeling compatibility rather than similarity. While similarity algorithms find items that look like each other, compatibility algorithms find items that go well with each other.
This distinction is critical in domains like fashion e-commerce. As explored by Mohamed Foued Ayedi et al. in Recommendation System for Enhanced Customer Experience: A Novel Image-to-Text Method and Mingzhe Yu et al. in Smart Fitting Room: A One-stop Framework for Matching-aware Virtual Try-On, a truly useful system must move past suggesting visually similar garments and instead understand how different pieces of clothing combine to form a visually coherent outfit. This concept extends beyond retail; for instance, Lina Wang et al. in PDSR: A Privacy-Preserving Diversified Service Recommendation Method on Distributed Data apply similar logic to recommend combinations of interdependent web services that fulfill complex, multi-stage workflows.
The Combinatorial Challenge of Set Optimization
Moving from items to sets transforms recommendation from a simple sorting problem into a complex combinatorial search. If a catalog contains thousands of items, the number of potential multi-item combinations explodes exponentially. Researchers have developed several distinct mathematical frameworks to make this search tractable:
1. Graph and Clique Optimization
If we represent items as nodes in a graph and draw edges between compatible items, finding a perfectly cohesive set of a fixed size $K$ can be modeled as finding a clique (a fully connected subgraph).
In Exact-K Recommendation via Maximal Clique Optimization, Yu Gong et al. applied this intuition to user interfaces that display items in fixed-size grids or "cards." Instead of filling the card with the top-$K$ independently ranked items, they used a Graph Attention Network (GAttN) trained via reinforcement learning to find a maximal clique, ensuring that every item on the card contextually reinforces the others.
2. Constrained Knapsack Optimization
When bundles must respect real-world limitations like a user's budget, the problem mirrors the classic knapsack optimization problem. Mohammad Khabbaz and Min Xie 0002 addressed this in TopRecs+: Pushing the Envelope on Recommender Systems by designing a two-phase algorithm that filters candidate items based on collaborative preferences and then uses knapsack-style optimization to assemble bundles (such as travel itineraries) that maximize total utility without exceeding the user's budget.
3. Joint Slate and Distribution Optimization
Even in traditional two-stage recommenders (which first nominate candidates and then rank them), treating stages or items in isolation causes issues. Jiri Hron et al. showed in On Component Interactions in Two-Stage Recommender Systems that failing to account for list-wide interactions during training leads to suboptimal "slates" (the final group of served items).
Similarly, Yiding Ran et al. demonstrated in PM K-LightGCN: Optimizing for Accuracy and Popularity Match in Course Recommendation that macro-properties, such as whether a user prefers a mix of popular and niche items, should be optimized across the entire set simultaneously rather than filtering items one by one.
Concurrent Set Completion and Global Representation
A major application of this paradigm is set completion: given a partial collection of items, how does the system fill in the remaining blanks?
This task differs fundamentally from sequential next-item prediction because it operates within a closed, finite boundary. As Iacopo Vagliano et al. observed in Recommendations for item set completion: on the semantics of item co-occurrence with data sparsity, input size, and input modalities, completing a finite set (like a list of citations or subject labels) requires understanding the boundaries of a completed group rather than just predicting a chronological sequence. To handle the massive data sparsity that occurs when treating groups as single entities, early approaches like the one by Advaith Siddharthan et al. in Matrix Factorization for Package Recommendations adapted collaborative filtering to learn latent factors for item combinations.
Sequential Prediction (Open-ended):
[Item 1] -> [Item 2] -> [Item 3] -> Predict [Next Item?]
Set Completion (Closed Boundary):
{ Item 1, Item 2, [ ? ], [ ? ] } -> Predict Concurrent Missing Components
Modern architectures solve this by learning holistic, set-level representations that allow for the simultaneous retrieval of multiple missing items. For instance, Yanan Cao et al. introduced S2SRec2: Set-to-Set Recommendation for Basket Completion with Recipe to predict multiple missing ingredients in a grocery basket at once, ensuring they collectively match a target recipe.
To achieve this at scale, researchers have turned to self-attention mechanisms. In OutfitTransformer: Learning Outfit Representations for Fashion Recommendation, Alan Lu et al. bypassed the limitations of pairwise item comparisons by using a Transformer to encode an entire collection into a single outfit-level embedding. Because self-attention computes relationships between all items in the input simultaneously, the model can look at a partial outfit and execute a single, concurrent multi-item retrieval query to find the most compatible remaining pieces.
Taking this a step further, Wang-Cheng Kang et al. showed in Visually-Aware Fashion Recommendation and Design with Generative Image Models that these holistic representations can even be hooked up to generative models. Instead of merely retrieving existing items to complete a set, the system can synthesize entirely new items tailored to fit the exact aesthetic gap in a user's collection.
Key Foundational Work
Breaking out of the box of recommendations: from items to packages (2010) — contributed 2 concept elements to this direction.
Learning user preferences for sets of objects (2006) — contributed 1 concept elements to this direction.