It can be viewed as a problem of classification, clustering, and topic distribution.

First classify by business, then search more carefully

Hidden Markov, conditional random fields, support vector machines, maximum entropy

Each category corresponds to several fixed text fields, such as title and category name.

Replacement: synonym replacement, abbreviation replacement

Delete: word selection (sequence labeling) word weight (TF-IDF)

Extensions: semantic (LDA, Word Embedding) terms, link analysis, session mining

Alignment: Machine Translation

session mining

Calculate weights: TF-IDF, rule scoring, statistical learning

Data collection: manual annotation, user click co-occurrence

Recall by category

Optimize sorting by category preference of query terms

Filter by category preference of query terms

Pre-search guidance: Mainly for recommendations

Guidance during search: mainly for completion

Post-search guidance: mainly to refine related query terms and merchant recommendations

Search guided traffic proportion, click-through rate, conversion rate

Data sources: search logs, WIFI information, geographical location, hotspot information

Data preprocessing: process raw data and generate recommended candidate words and hot words

Recall strategy layer: collaborative filtering, association rules, retrieval, click bipartite graph, etc.

Ranking layer: LTR, progressive gradient boosting tree. Rule sorting-learning sorting-post-processing

Display layer: manual intervention, organizational display

Data sources: logs, business data, locations, categories, industry terms, combined words

Algorithm: Trie tree, HashTable, inverted index; ordinary prefix, Pinyin prefix, Simplified Pinyin prefix

It is necessary to consider the blocked area when the user is typing and consider the effective display area.

Insert ad slot at fixed location