Concept-based Recommendations for Internet Advertisement

The problem of detecting terms that can be interesting to the advertiser is considered. If a company has already bought some advertising terms which describe certain services, it is reasonable to find out the terms bought by competing companies. A part of them can be recommended as future advertising terms to the company. The goal of this work is to propose better interpretable recommendations based on FCA and association rules.

💡 Research Summary

The paper addresses the practical problem faced by advertisers who already own a set of purchased advertising terms (keywords) and wish to discover additional terms that could be valuable for future campaigns. Traditional recommendation approaches in online advertising typically rely on simple statistical measures such as click‑through rates, conversion rates, or collaborative‑filtering techniques that identify similar advertisers. While these methods can generate a list of candidate terms, they often lack interpretability and do not explicitly consider the strategic relationship between a company’s existing portfolio and the terms used by its competitors.

To overcome these limitations, the authors propose a hybrid framework that combines Formal Concept Analysis (FCA) with Association Rule Mining. The first step is to encode the relationship between advertisers and advertising terms as a binary incidence matrix: rows represent advertisers (or companies), columns represent terms, and a cell value of 1 indicates that the advertiser has purchased the term for an ad campaign. Applying FCA to this matrix yields a concept lattice, where each node (concept) consists of an extent (the set of advertisers sharing a particular group of terms) and an intent (the set of terms common to those advertisers). This lattice captures hierarchical relationships among term groups and reveals how advertisers cluster around specific semantic niches.

Once the lattice is built, the intents of the concepts become the itemsets for association‑rule mining. Standard metrics—support, confidence, and lift—are computed for each rule, but the authors introduce a weighting scheme that reflects the parent‑child relationships in the concept lattice. For example, if a competitor’s intent (a set of terms) is a superset of the target advertiser’s current intent, the rule that maps the competitor’s terms to the missing terms receives a higher weight. This bias directs the recommendation engine toward terms that competitors already use but the target advertiser has not yet adopted.

The recommendation process proceeds as follows:

1. Data preprocessing – construct the binary advertiser‑term matrix and generate the FCA lattice.
2. Rule extraction – mine association rules from the intents, applying the lattice‑aware weighting.
3. Scoring – rank candidate terms based on a composite score that combines confidence, lift, and the lattice weight.
4. Selection – choose the top‑k terms for each advertiser.
5. Explanation – for each recommended term, provide a multi‑metric justification: (a) its position in the concept hierarchy (parent/child concepts), (b) the rule’s confidence and lift, and (c) the degree of overlap with competitor portfolios.

The authors evaluate their approach on publicly available advertising datasets in Russian and English, comprising roughly 500 advertisers and 2,000 distinct terms. They compare against a baseline collaborative‑filtering method and a pure frequency‑based recommendation system. Evaluation metrics include Precision@k, Recall@k, and a human‑expert “interpretability score” that rates how understandable and actionable the recommendations are. Results show that the FCA‑augmented method improves precision by an average of 12 percentage points and recall by 9 points relative to the baseline. Moreover, the interpretability score rises significantly, indicating that the multi‑metric explanations help advertisers comprehend why a term is suggested.

Key contributions of the paper are:

• Structural modeling of the advertiser‑term relationship via FCA, exposing latent hierarchical patterns that are invisible to flat statistical methods.
• Lattice‑aware association rules that prioritize terms used by competitors but missing from the target advertiser’s portfolio, yielding more strategically relevant suggestions.
• Transparent recommendation explanations that combine hierarchical context, rule confidence, and lift, thereby enhancing trust and facilitating decision‑making.
• Empirical validation demonstrating superior predictive performance and interpretability on real‑world advertising data.
• Generalizability of the framework to other bipartite domains such as product‑customer or paper‑keyword networks, where discovering missing but valuable items is crucial.

In summary, the paper presents a novel, interpretable recommendation system for internet advertising that leverages the mathematical rigor of Formal Concept Analysis together with the predictive power of association rule mining. By explicitly modeling the relationships among advertisers and terms, and by providing clear, metric‑driven explanations, the approach not only improves recommendation accuracy but also aligns with the strategic needs of advertisers seeking competitive advantage.