Core Distinctions:

• Advertising: When a company builds an advertising system, the commercial goal is straightforward – address the revenue generation. Hence, the algorithms in advertising aim directly at boosting the company’s revenue.

• Recommendation: While the end goal of recommendation algorithms is also revenue augmentation, the immediate objective is enhancing user engagement. Increased engagement paves the way for more advertising inventory, subsequently driving up the revenue.

• Search: The crux of search algorithms revolves around the user-inputted search terms. Despite the rising emphasis on personalized results, the essence of personalization in recommendation algorithms serves only as a supplement to search algorithms. Efficiently fetching information around the search terms remains the fundamental objective of search algorithms.

Optimization Objectives Divergence:

• Advertising: Unified across companies, the estimation goals for advertising algorithms are the Click-Through Rate (CTR) and Conversion Rate (CVR), closely tied to the current product form of performance-based advertising.

• Recommendation: The estimation objectives here diverge, with video platforms leaning towards predicting watch duration, news platforms predicting CTR, and e-commerce platforms estimating average order value, all of which relate closely to user engagement.

• Search: The estimation objectives in search vary again since, unlike advertising and recommendation, search inherently has a “correct answer” to a certain extent, emphasizing on recalling these correct answers.

Differences in Algorithm Model Design:

• Advertising: Precision in predicting CTR and CVR for accurate bidding is crucial, leading to rigorous calibration requirements in advertising algorithms.

• Recommendation: The results are often presented as a list, demanding a broader view, often encompassing list-level or page-level algorithms to optimize the overall effect.

• Search: The emphasis is on understanding search terms and the content, necessitating the heavy application of NLP models for better semantic understanding of user queries.

Explorations and Exploitations in Recommendation Systems:

The Pain Points at System Level:

• Advertising Algorithms: Achieving global profit maximization through the synergistic operation of various modules is a monumental challenge, with the complexity often reaching unmanageable levels.

• Recommendation Algorithms: The eternal struggle lies in balancing short-term and long-term benefits, a puzzle every recommendation algorithm engineer dreams of solving.

• Search Algorithms: The focus is placed heavily on understanding the content of search terms and items, with the comprehension of multimodal content like images and videos often being the bottleneck in improving search effectiveness.