Викторина, связанная с руководством · Средний Уровень
Учимся ранжироваться в викторине по поиску товаров
Тестирование LTR при поиске продуктов: точечное, парное и списочное обучение; метки релевантности; НДЦГ; смещение щелчка; жесткие ограничения; и онлайн-ограждения.
Вопрос 1 из 8
What does an LTR model produce for a query?
Предварительный просмотр каждого вопроса8 вопросы
Ответы и пояснения появляются по мере прохождения теста.
- What does an LTR model produce for a query?
- A single product category for each catalog item
- An ordered list of candidate products based on learned relevance signals
- A new product description
- A legal guarantee that the first item is best
- How does pairwise LTR frame training examples?
- As a single label for the whole catalog
- As an image segmentation mask
- As a list of all possible product categories
- As preferences indicating which of two items should rank higher
- Why can click data be biased as relevance labels?
- Clicks depend on exposure and position as well as product relevance
- Every product receives equal exposure
- A click measures price accuracy only
- Click logs contain no user behavior
- A shopper selects size medium. How should a ranker handle items available only in large?
- Keep them if their margin is high
- Lower their score but leave them as equivalent options
- Filter them out before ranking results for that constraint
- Let popularity override size
- Which metric discounts relevance lower in the search-result list?
- БЛЮ
- NDCG
- Частота ошибок в словах
- Mean pixel error
- Why should the test split hold out queries or respect time?
- To reduce leakage from near-duplicate queries or future interactions into training
- To make the model score every item identically
- To avoid measuring relevance
- To increase the number of training labels
- What can a reranker not do if a relevant product was never retrieved?
- Use a price feature
- Reorder candidates
- Apply a relevance score
- Promote a candidate that is absent from the candidate set
- Which safeguard helps reveal the effect of a business feature such as margin?
- Add the feature without logging it
- Optimize margin alone
- Run an ablation or controlled comparison and monitor relevance and exposure guardrails
- Remove all human judgments