Back

Tengence Recommendation System: Delivering Personalized Recommendations at Scale

Tengence Team ·
Technical Solutions Recommendation System E-commerce
Tengence Recommendation System: Delivering Personalized Recommendations at Scale

Summary: As catalogs, content volume and customer touchpoints keep expanding, e-commerce platforms are no longer asking only how to win traffic, but how to help each shopper reach the right product faster. This article explains what the Tengence recommendation system does, how the recall-to-rerank pipeline works, how it is integrated, and what a three-month optimization project delivered.

Introduction

Product counts, content libraries and customer touchpoints keep growing, and the core question for e-commerce platforms has shifted from "how do we get traffic" to "how do we help each shopper find relevant products faster". When a recommendation system still leans mainly on bestsellers and basic collaborative filtering, the same problems tend to show up: user profiles that are too coarse, recommendations that look identical to everyone, behavior feedback that arrives too late, and limited adaptability across business scenarios.

Before optimization, one fashion apparel platform reported the following core metrics:

Core metric Current value Industry benchmark Gap
Recommendation CTR 2.1% 3.5%+ -40%
Recommendation CVR 0.8% 1.5%+ -47%
Recommendation GMV contribution 12% 25%+ -52%
User dwell time 1 min 28 s 3 minutes+ -51%
Recommendation coverage 35% 70%+ -50%

These numbers show that a recommendation system has to do more than rank items more accurately. It also has to solve data collection, feature engineering, real-time computation, multi-scenario adaptation and continuous experimentation. This article walks through the business value, algorithm capabilities, technical architecture, integration paths, a case implementation, and practical optimization methods.

1. Why a Recommendation System Is Core Infrastructure for E-Commerce

1.1 Information overload raises the cost of choice

As products and content keep accumulating, shoppers have to filter and decide within a limited amount of time. A recommendation system uses user behavior, item content and context signals to surface more relevant candidates, which lowers the cost of choice and improves how efficiently platform traffic is used.

Shopper-side problem Typical symptom Impact on the platform
High cost of choice Too many products to compare quickly Many products get little effective exposure
Limited attention Shoppers cannot browse and filter for long Traffic conversion efficiency drops
Unclear intent Shoppers often arrive without a specific purchase goal Demand has to be stimulated through content matching and scenario-based recommendations

The value of a recommendation system is not simply showing more products. It is dynamically adjusting relevance, ordering and combination based on a shopper's current intent and long-term preferences.

1.2 Public industry cases show the commercial value of recommendations

Public industry cases usually link recommendation systems to revenue contribution, content discovery and user time spent. Business models, measurement definitions and recommendation scenarios differ from platform to platform, so the figures below are best read as industry references rather than as fixed outcomes for every platform.

Platform Value of the recommendation system Publicly reported data
Amazon Revenue contribution from product recommendations About 35% of revenue comes from product recommendations
Netflix More efficient content discovery More than 80% of content is discovered through personalized recommendations
Content platforms Better content matching and longer session time Recommendation algorithms are core infrastructure for content distribution

For e-commerce platforms operating at scale, recommendation has grown from a single marketing feature into core infrastructure that connects user demand, product supply and business goals.

1.3 Business metrics for personalized recommendations

Based on the industry case data cited above, personalized recommendations may move the following metrics. Actual results depend on data quality, business scenario, traffic mix, model capability and experiment design.

Metric Without recommendations With recommendations Lift
Click-through rate 2.8% 4.5% +61%
Conversion rate 1.9% 3.2% +68%
Average order value ¥128 ¥156 +22%
Dwell time 1 min 28 s 3 min 12 s +127%
GMV contribution 12% 28% +133%

When evaluating a recommendation system, look at short-term clicks, transaction conversion, user experience and long-term user value together, rather than judging the system on a single metric.

2. Core Capabilities of the Tengence Recommendation System

2.1 Four families of algorithms

The Tengence recommendation system combines collaborative filtering, content-based recommendation, deep learning and hybrid recommendation based on data volume, business goals and scenario constraints.

Collaborative filtering

Algorithm type Description Best-fit scenario Main value
User-based collaborative filtering Finds similar users from behavior data and recommends products those users prefer Scenarios with rich user behavior data Improves discovery and supports new-product exposure
Item-based collaborative filtering Finds similar products from a user's historical behavior Scenarios where product relationships are relatively stable Improves relevance; works well for related recommendations
Matrix factorization Uses latent factor models to mine hidden relationships between users and items Large-scale recommendation scenarios Balances compute efficiency against recommendation quality
Cold-start optimization Uses item content and basic attributes to compensate for sparse history New-user and new-product scenarios Raises initial recommendation coverage

Content-based recommendation

Algorithm type Description Best-fit scenario Main value
Content feature matching Matches user preferences against product attributes and content features Scenarios with a mature tagging system Strong interpretability
Tag similarity Computes item similarity from category, brand and tags Scenarios with a clear taxonomy Supports related-product recommendations
Text semantic analysis Uses NLP on titles, descriptions and detail-page text Scenarios with rich text information Improves semantic matching
Image feature extraction Uses CV to extract features from product images Apparel, home and other visually driven categories Adds visual similarity to the signal mix

Deep learning recommendation

Algorithm type Description Best-fit scenario Main value
Deep neural network models Learns complex relationships between users, items and context Scenarios with large data volumes Better modeling of complex features
Sequential recommendation Models interest drift from behavior sequences Scenarios where intent is time-dependent Captures short-term interest changes
Online learning Updates models or features as the data distribution shifts Scenarios with fast-changing preferences Shortens the behavior feedback loop
Multi-objective optimization Considers CTR, CVR and average order value together Scenarios that must balance several business metrics Serves both short-term conversion and overall revenue

Hybrid recommendation

Capability type Description Best-fit scenario Main value
Multi-algorithm fusion Blends results from several recall and ranking paths Scenarios needing both coverage and relevance Exploits the strengths of different algorithms
Rule engine integration Folds promotion, inventory and block rules into the recommendation path Scenarios with complex business rules Improves controllability and execution flexibility
Business policy configuration Adjusts recommendation strategy and weights through configuration Scenarios where operating strategy changes often Shortens the policy adjustment cycle
A/B testing support Runs experiments across models, features or rules Scenarios that need continuous optimization Lets data decide between options

2.2 Real-time personalization

A recommendation system has to separate three things: real-time feature updates, real-time recommendation computation, and model updates. User behavior can enter the feature pipeline quickly, while model updates usually follow an independent training and release cycle.

Real-time feature updates:

Feature type Update frequency Update method Business value
User profile Real time Updated within 100ms of a user action Reflects the latest interest shifts
Behavior features Immediate Computed as interaction events arrive Captures current visit intent
Interest preferences Dynamic Continuously adjusted from the behavior sequence Adapts to changing user interests

Real-time recommendation computation:

Capability Metric Description
Response time Under 100ms on average Millisecond-level recommendation responses served from the backend
Latency target Under 200ms for 99.9% of requests Constrains end-to-end latency in the recommendation path
Online inference Real-time model inference Returns recommendations based on request context
Cache optimization Multi-level caching strategy Shortens the access path for hot data

2.3 Multi-scenario adaptation

The same recommendation system has to adjust its ranking logic to the goal of each page instead of reusing one set of results everywhere.

Scenario Recommendation strategy Business goal Main capability
Storefront homepage Balances user preference, popularity and content diversity Balancing personalization with discovery Hybrid recommendation
Product detail page Recommends related content from the current product and user preferences Supporting the purchase decision Collaborative filtering + content recommendation
Shopping cart Recommends complementary products based on purchased or pending items Raising average order value Association rules + personalization
Account center Personalized ordering based on long-term preferences Driving repeat visits and clicks Deep learning recommendation
Campaign page Combines campaign rules with user attributes Raising campaign participation and conversion Rule engine + personalization
Search results page Blends search intent with user preferences Improving search result relevance Search-recommendation fusion

3. How a Recommendation System Works: A Deep Dive

3.1 Data collection: the foundation

User behavior data:

Data type Content Collection method Update frequency
Browsing behavior Product views, page dwell time, scroll depth Event tracking Real time or near real time
Click behavior Clicks on recommendation slots and product cards Event tracking Real time
Favorite behavior Favoriting products and stores Event collection Real time
Add-to-cart behavior Adding to cart, changing quantity Event collection Real time
Purchase behavior Order placement, payment, refunds Order synchronization Synced along the business pipeline
Search behavior Search terms, filter conditions Search logs Real time or near real time
Review behavior Product reviews and ratings Review system Near real time

Item data:

Data type Content Source
Product attributes Category, brand, material, size, color Product catalog
Content features Title, description, detail-page text Product catalog
Image features Main image, detail images Product catalog + CV extraction
Statistical features Click-through rate, conversion rate, repeat-purchase rate Behavior statistics
Relationship features Similar products, complementary products Relationship computation

Context data:

Data type Content Collection method
Time context Current time, day of week, time band Client
Location context Geographic location, IP origin Client / IP
Device context Device type, operating system, browser Client
Channel context Source channel, UTM parameters URL parameters
Campaign context Active campaigns, promotion information Campaign system

3.2 Feature engineering: the step that decides how well recommendations perform

User feature system:

Feature group Specific features Feature type
Basic attributes Age, gender, region, tenure since registration Static features
Interest preferences Preferred categories, brands and price bands Statistical features
Behavior features Activity level, purchasing power, loyalty, price sensitivity Statistical features
Real-time state Latest views, current session, recent purchases Real-time features
Lifecycle stage New, growing, mature, churned Categorical features

Item feature system:

Feature group Specific features Feature type
Content features Text embedding, image embedding Deep features
Attribute features Category, brand, price, tags Categorical / numeric features
Statistical features CTR, CVR, impressions, sales volume Statistical features
Relationship features Similar-item list, complementary-item list Relationship features
Quality features Positive review rate, return rate, complaint rate Statistical features

The goal of feature engineering is not simply to add more features. It is to help recall, ranking and business rules identify user intent, item value and context constraints more accurately.

3.3 Recommendation computation: recall, pre-ranking, ranking, re-ranking

Recommendation services usually adopt a four-stage architecture that balances candidate coverage, compute efficiency, ranking accuracy and business constraints.

Recall

The recall stage builds a candidate set from multiple sources, with the focus on coverage and candidate diversity.

Recall strategy Description Candidate count Main trait
Collaborative filtering recall Generates candidates from user-item relationships 500-1000 Strong discovery
Content recall Generates candidates from item content similarity 500-1000 Strong relevance
Popularity recall Generates candidates from overall or scenario heat 200-500 Provides a basic fallback
Interest recall Generates candidates from user interest tags 500-1000 Highly personalized
Geo recall Generates candidates from location and regional features 100-300 Good for scenario-based recommendations
Multi-channel recall fusion Merges sources and removes duplicates 2000-5000 Combines multiple signals

Pre-ranking

Pre-ranking quickly narrows several thousand candidates down to a few hundred, reducing the compute load on the ranking model.

Pre-ranking strategy Description Items processed Response time
Fast scoring model Uses a lightweight model for preliminary scoring 2000-5000 <10ms
Rule filtering Removes candidates blocked by inventory, status or business rules 2000-5000 <5ms
Deduplication Removes duplicate products or duplicate content 500-1000 <5ms

Ranking

The ranking stage scores the surviving candidates with a finer-grained model.

Ranking strategy Description Items processed Response time
Deep learning model Uses Wide&Deep, DeepFM, DIN and similar models 200-500 <50ms
Multi-objective optimization Considers CTR, CVR and average order value jointly 200-500 <30ms
Business rule integration Adds inventory, promotion and margin constraints 200-500 <10ms

Re-ranking

Re-ranking handles diversity, slots and business policy on top of the model output.

Re-ranking strategy Description Output count Response time
Diversity control Prevents over-concentration of similar products or content 50-100 <10ms
Business rule adjustment Applies pinning, blocking and policy ordering 50-100 <5ms
Slot optimization Adjusts ordering to the display slot 50-100 <5ms

Overall performance targets:

Performance metric Target Description
Average response time <100ms Average elapsed time from request to recommendation response
P99 response time <200ms Response time at the 99th percentile stays below this value
Availability >99.9% Availability target for the recommendation service
Throughput >100,000 QPS Capacity for high-concurrency recommendation requests

4. Deploying and Integrating the Recommendation System

4.1 API integration

// Initialize the recommendation client
const tengenceRec = new TengenceRecommendation({
  apiKey: 'your-api-key',
  sceneId: 'homepage',
  endpoint: 'https://rec.tengence.com/api/v1'
});

// Fetch personalized recommendations
const recommendations = await tengenceRec.getRecommendations({
  userId: 'user_12345',
  count: 10,
  context: {
    page: 'homepage',
    device: 'mobile',
    itemId: null // no item ID on the homepage; pass the current item ID on detail pages
  }
});

// Example response
// {
//   "code": 0,
//   "data": {
//     "recommendations": [
//       {"itemId": "1001", "score": 0.95, "reason": "Based on your browsing history"},
//       {"itemId": "1002", "score": 0.89, "reason": "Similar to items you favorited"}
//     ],
//     "traceId": "abc123",
//     "modelVersion": "v2.3.1"
//   }
// }

traceId supports request tracing and troubleshooting, and modelVersion identifies which model version served a request. In production, the client logic should be completed against your authentication, error handling and data compliance requirements.

4.2 SDK integration

Tengence provides SDKs for multiple languages and platforms:

SDK language Supported platforms Main traits
JavaScript/TypeScript Web, mini programs, Node.js Covers front-end and server-side scenarios
Java Android, server side Suited to enterprise application integration
Python Server side Convenient for data processing and model collaboration
PHP Web server side Fast integration for web businesses

4.3 Plugin integration

Platform Plugin type Installation
WordPress Recommendation plugin Install from the plugin directory
Shopify Recommendation app Install from the App Store
Magento Recommendation module Install via Composer

4.4 Behavior data reporting

// Report behavior data
tengenceRec.report({
  userId: 'user_12345',
  eventType: 'view', // view/click/fav/cart/purchase
  itemId: '1001',
  context: {
    page: 'homepage',
    position: 1,
    ref: 'recommendation'
  },
  timestamp: Date.now()
});

Reported data usually covers the user identifier, event type, item identifier, page scenario, display position and timestamp. Actual fields should be defined by your tracking plan and privacy compliance requirements.

5. Case Study: Three Months of Recommendation Optimization at an E-Commerce Platform

5.1 Background and problem diagnosis

The platform in this case is a fashion apparel e-commerce site founded in 2019, with annual GMV of about ¥180M, about 500,000 monthly active users and roughly 30,000 SKUs.

The main problems before optimization were:

Metric Before optimization Industry reference Gap
Recommendation CTR 2.1% 3.5%+ -40%
Recommendation CVR 0.8% 1.5%+ -47%
Recommendation GMV contribution 12% 25%+ -52%
User dwell time 1 min 28 s 3 minutes+ -51%

5.2 Solution and implementation timeline

The implementation ran for 8 weeks across four phases:

The technical architecture:

Data layer
  ├─ User behavior data (real-time collection)
  ├─ Item content data (synced from product catalog)
  └─ Context data (reported by the client)
        ↓
Feature layer
  ├─ User features (100+ dimensions)
  ├─ Item features (200+ dimensions)
  └─ Context features (50+ dimensions)
        ↓
Recall layer
  ├─ Collaborative filtering recall
  ├─ Content recall
  ├─ Popularity recall
  └─ Interest recall
        ↓
Pre-ranking layer
  ├─ Lightweight scoring model
  └─ Business rule filtering
        ↓
Ranking layer
  ├─ DeepFM model
  ├─ Multi-objective optimization (CTR + CVR)
  └─ Business rule integration
        ↓
Re-ranking layer
  ├─ Diversity control
  ├─ Business rule adjustment
  └─ Slot optimization
        ↓
Recommendation output

5.3 Month 1: data onboarding and feature engineering

Week Task Completion Key result Risk and response
Week 1 Data onboarding design 100% Data collection plan confirmed —
Week 2 User behavior data onboarding 100% Historical data migrated Insufficient data volume; widened collection scope
Week 3 Item data onboarding 100% Product feature engineering completed Some tags missing; built a backfill mechanism
Week 4 Feature engineering development 100% 350+ features in production —

5.4 Month 2: model training and canary preparation

Week Task Completion Key result Risk and response
Week 5 Recall algorithm development 100% 4 recall channels live —
Week 6 Pre-ranking and ranking model training 100% DeepFM model trained Insufficient training data; applied data augmentation
Week 7 A/B test preparation 100% Experiment plan completed —
Week 8 Canary launch 100% 10% of traffic in the experiment —

5.5 Month 3: canary expansion and full rollout

Week Task Completion Key result Risk and response
Week 9 Canary expanded to 30% 100% Recommendation CTR up 35%, CVR up 28% —
Week 10 Canary expanded to 50% 100% Overall effect stayed positive —
Week 11 Canary expanded to 100% 100% Full switchover completed Fast response when monitoring flagged anomalies
Week 12 Continuous optimization 100% Recommendation performance improving steadily —

5.6 Results after three months

Core metric changes:

Core metric Before After Lift Industry comparison
Recommendation CTR 2.1% 3.8% +81% Above the industry reference value
Recommendation CVR 0.8% 1.5% +88% Reached the industry reference value
Recommendation GMV contribution 12% 28% +133% Above the industry reference value
User dwell time 1 min 28 s 3 min 5 s +119% Above the industry reference value
Average order value ¥128 ¥152 +19% Still improving

Results by scenario:

Scenario CTR lift CVR lift GMV contribution lift
Homepage recommendations +75% +82% +125%
Detail-page recommendations +88% +95% +145%
Cart recommendations +95% +78% +110%
Account-center recommendations +70% +85% +130%

ROI analysis:

Investment item Amount (RMB, in units of ten thousand yuan) Return item Amount (RMB, in units of ten thousand yuan)
System build-out 25 GMV growth 280
Data onboarding 8 Lower customer acquisition cost 35
Operations optimization 5 Higher existing-user value 45
Total investment 38 Total return 360

Combining the ¥380,000 invested with ¥3,600,000 in returns, the return on investment in this case is roughly 1:9.5. These figures are the staged result of an anonymized case over a three-month period; real projects should be evaluated independently against traffic scale, business structure and experiment definitions.

6. Recommendation Optimization Practices

6.1 Cold-start optimization

Cold start has to be handled separately for new users and new products.

New-user cold-start strategies:

Strategy Description Effect
Popularity fallback Shows scenario bestsellers in the initial phase Guarantees a baseline recommendation experience
Registration guidance Prompts users to fill in basic preferences Builds an initial profile quickly
Preference questionnaire Collects interest signals through 3-5 questions Shortens the time to personalization
Real-time behavior learning Updates features after every action Gradually forms a dynamic profile

New-product cold-start strategies:

Strategy Description Effect
Content-based recommendation Uses product attributes, text and image features Provides initial recommendations without historical behavior
Similar-product exploration Finds related products by content similarity Earns initial exposure
Explore-exploit strategy Balances exploring new products against exploiting known preferences Serves both short-term results and long-term returns
Small-traffic test exposure Starts with limited exposure, then adjusts traffic based on data Reduces the risk of recommending new products

6.2 Multi-objective optimization

Recommendation systems usually have to balance several objectives at once:

Objective Optimization strategy Weighting considerations
Click-through rate Interest matching, title and image optimization Tuned to traffic value
Conversion rate Purchase intent recognition, price sensitivity analysis Tuned to margin structure
Average order value Weighting high-value products, bundle recommendations Tuned to business goals
Long-term value User satisfaction, repeat-purchase rate, LTV Tuned to long-term strategy

One way to express multi-objective ranking is:

Final score = w1 x pCTR + w2 x pCVR + w3 x price factor + w4 x diversity factor

The specific weights and factor definitions should be configured against the platform's margin structure, inventory policy, user experience and current-stage goals; they cannot be fixed independently of the business scenario.

6.3 Continuous optimization through A/B testing

Test dimensions:

Dimension Test content Reference cycle
Algorithm choice Collaborative filtering versus deep learning models 2-4 weeks
Parameter tuning Learning rate, regularization parameters and so on 1-2 weeks
Feature engineering New features, feature combinations 2-3 weeks
Business rules Diversity weights, policy rules 1-2 weeks

Evaluation metrics:

Metric type Specific metrics Evaluation focus
Click metrics CTR, click depth Content relevance and appeal
Conversion metrics CVR, AOV, GMV Transaction contribution and business outcomes
Experience metrics Dwell time, bounce rate User experience and content fit
Satisfaction metrics Positive review rate, repeat-purchase rate Long-term user value

Experiments should keep a stable control group and treatment group, and compare short-term conversion against long-term user value over a long enough observation window, rather than adjusting models or rules on a single fluctuation.

7. FAQ

Q: How much data do we need before using a recommendation system?

A: Tengence supports cold start, so a small amount of data is enough to build a baseline recommendation capability.

Minimum data requirements for reference:

As data volume grows, models usually pick up richer user-item relationships and context signals:

Actual data requirements still depend on the business scenario, catalog size, behavior distribution and data quality.

Q: What is the response time of the recommendation system?

A: Current performance references:

Performance metric Target Reference actual value
Average response time <100ms About 60ms
P99 response time <200ms About 150ms
Availability >99.9% 99.95%
Throughput >100,000 QPS Supports 150,000 QPS

P99 means the 99th-percentile request latency, that is, roughly 99% of requests respond within that time. Actual performance depends on request volume, number of recall channels, model complexity, network conditions and deployment topology.

Q: How should we measure recommendation performance?

A: We recommend evaluating across four metric families:

Metric type Core metrics Evaluation method
Click metrics CTR, click depth Compared through A/B tests
Conversion metrics CVR, AOV, GMV Compared through A/B tests and attribution analysis
Experience metrics Dwell time, bounce rate Assessed through behavior data analysis
Satisfaction metrics Positive review rate, repeat-purchase rate Assessed through user surveys and long-term cohort analysis

In practice you should also watch recommendation coverage, diversity, inventory consumption, return rate and long-term user value.

Q: Does the recommendation system support real-time updates?

A: Real-time feature updates and real-time recommendation computation are supported; model updates follow an independent training cycle:

Capability Metric Description
Real-time feature updates Within 100ms Relevant features update within 100ms of a user action
Real-time recommendation computation <100ms Recommendation requests return within the target time
Model updates Near real time, hourly Models update based on data distribution and training policy

Real-world real-time performance depends on the whole chain spanning data reporting, feature processing, model serving and business systems.

8. Conclusions and Next Steps

8.1 Core lessons

Building a recommendation system requires thinking about algorithms, data, engineering and business policy together:

  1. Clear business goals: watch clicks, conversion, GMV, user experience and long-term value at the same time.
  2. Clean layered architecture: use recall, pre-ranking, ranking and re-ranking to balance coverage, accuracy and response speed step by step.
  3. Effective real-time feedback: ingest user behavior quickly to shorten the loop between an interest shift and the recommendation result.
  4. Differentiated scenario strategy: tune recommendation logic to the goals of the homepage, detail page, cart, account center and search page.
  5. Continuous experimentation: keep optimizing models, features and business rules through A/B testing and metric monitoring.

8.2 Implementation recommendations

When building or optimizing a recommendation system, companies can follow this path:

Timeframe Action Expected result
This week Assess current recommendation performance and set optimization goals Clarify problems and metric baselines
Next week Choose a recommendation approach that matches business scale Define the data and implementation plan
Within 1-2 months Complete data onboarding, feature engineering and model training Basic recommendation capability live
Within 2-3 months Run canary experiments and observe results Initial CTR, CVR and GMV results
Ongoing Run A/B tests and tune strategy Better long-term recommendation performance

The timelines above are typical references; real schedules depend on data readiness, system integration, the number of business scenarios and experiment requirements.

8.3 What Tengence provides

Tengence delivers an AI recommendation system solution for enterprise businesses:

By understanding user behavior, matching content and ranking in real time, a recommendation system can deliver more relevant and more controllable results across business scenarios.

About Tengence

The recall-ranking-reranking architecture behind recommendation systems is the technical foundation of enterprise growth, and it belongs to the growth technology field Tengence focuses on. Tengence (TENGENCE) is the enterprise growth engine for the AI era, focused on public-traffic acquisition and private-domain conversion.

Related Reading

Data Sources

Get Started

Get Started
Book a free expert consultation and diagnostic

← Previous: Tengence Search Optimization: 5 Keys to a Better Site Search Experience Next: Unified Search and Recommendations: From Integrated Data to Coordinated Optimization →
Consult
QR code

WeCom